CN111753773A - Surface covering object recognition method, and neural network training method and device - Google Patents
Surface covering object recognition method, and neural network training method and device Download PDFInfo
- Publication number
- CN111753773A CN111753773A CN202010606948.7A CN202010606948A CN111753773A CN 111753773 A CN111753773 A CN 111753773A CN 202010606948 A CN202010606948 A CN 202010606948A CN 111753773 A CN111753773 A CN 111753773A
- Authority
- CN
- China
- Prior art keywords
- pixel point
- neural network
- label information
- sample image
- earth surface
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Withdrawn
Links
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 250
- 238000000034 method Methods 0.000 title claims abstract description 73
- 238000012549 training Methods 0.000 title claims abstract description 54
- 238000012545 processing Methods 0.000 claims description 15
- 238000004590 computer program Methods 0.000 claims description 9
- 230000006870 function Effects 0.000 description 32
- 238000010586 diagram Methods 0.000 description 9
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 description 7
- 239000003086 colorant Substances 0.000 description 6
- 238000013527 convolutional neural network Methods 0.000 description 6
- 230000008569 process Effects 0.000 description 6
- 244000025254 Cannabis sativa Species 0.000 description 4
- 230000008901 benefit Effects 0.000 description 3
- 238000004891 communication Methods 0.000 description 3
- 230000008878 coupling Effects 0.000 description 3
- 238000010168 coupling process Methods 0.000 description 3
- 238000005859 coupling reaction Methods 0.000 description 3
- 230000008859 change Effects 0.000 description 2
- 238000012512 characterization method Methods 0.000 description 2
- 238000013507 mapping Methods 0.000 description 2
- 238000005070 sampling Methods 0.000 description 2
- 238000012216 screening Methods 0.000 description 2
- 238000011161 development Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000004088 simulation Methods 0.000 description 1
- 239000002689 soil Substances 0.000 description 1
- 230000003595 spectral effect Effects 0.000 description 1
- 238000001228 spectrum Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
- G06V20/13—Satellite images
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Biophysics (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- General Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Biomedical Technology (AREA)
- Health & Medical Sciences (AREA)
- Astronomy & Astrophysics (AREA)
- Remote Sensing (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The present disclosure provides a surface covering identification method, a neural network training method and a device, wherein the surface covering identification method comprises: acquiring a ground surface image; carrying out classification and identification of pixel point levels on the earth surface image by using a target neural network to obtain the category information of an earth surface covering to which the pixel points in the earth surface image belong; the target neural network is obtained by training through a surface sample image and target label information corresponding to sample pixel points in the surface sample image, the target label information is used for identifying the surface covering object type to which the sample pixel points belong, and the target label information is obtained by adjusting initial label information corresponding to the sample pixel points in the surface sample image.
Description
Technical Field
The present disclosure relates to the field of computer vision technologies, and in particular, to a method for identifying a surface covering, a method and an apparatus for training a neural network, an electronic device, and a storage medium.
Background
The ground surface covering is a comprehensive body of a plurality of elements of the ground surface covered by natural buildings and artificial buildings on the surface of the earth, and comprises ground surface vegetation, soil, glaciers, rivers, lakes, swamp wetlands and various buildings, and the ground surface covering and the change are important scientific data and key parameters which are indispensable in the aspects of environmental change research, earth system mode simulation, geographical national condition monitoring, sustainable development planning and the like.
Generally, the type of the ground surface covering can be acquired by high-precision satellite image data or manual measurement, and a high-precision ground surface covering image is constructed based on the acquired type of the ground surface covering.
Therefore, how to effectively obtain the category of the ground cover so as to construct a high-precision ground cover image is still a problem which needs to be solved at present.
Disclosure of Invention
The disclosed embodiments provide at least one surface covering identification scheme.
In a first aspect, an embodiment of the present disclosure provides a method for identifying a surface covering, including:
acquiring a ground surface image; carrying out classification and identification of pixel point levels on the earth surface image by using a target neural network to obtain the category information of an earth surface covering to which the pixel points in the earth surface image belong; the target neural network is obtained by training through a surface sample image and target label information corresponding to sample pixel points in the surface sample image, the target label information is used for identifying the surface covering object type to which the sample pixel points belong, and the target label information is obtained by adjusting initial label information corresponding to the sample pixel points in the surface sample image.
The method and the device can be used for carrying out classification and identification of pixel point levels on the earth surface image by utilizing the pre-trained target neural network, in addition, the target neural network is obtained by training based on target label information corresponding to the sample pixel points in the earth surface sample image in a training stage, and the target label information obtained by utilizing initial label information adjustment is more accurate, so that the identification accuracy of the earth surface covering can be improved when the classification information of the earth surface covering to which each pixel point in the earth surface image belongs is predicted based on the target neural network.
In a possible implementation manner, after obtaining the category information of the surface covering to which the pixel points in the surface image belong, the method for identifying the surface covering further includes: determining the display color of the pixel point in the earth surface covering image corresponding to the earth surface image based on the category information corresponding to each pixel point in the earth surface image; and generating a ground surface covering image corresponding to the ground surface image based on the display color corresponding to each pixel point.
In the embodiment, the display colors of the pixel points are respectively determined according to the category information corresponding to the pixel points, so that the areas where the earth surface coverings of different categories are located can be clearly identified in the earth surface chart.
In a second aspect, an embodiment of the present disclosure provides a training method for a neural network, including:
acquiring a ground surface sample image and an initial ground surface coverage image corresponding to the ground surface sample image, wherein each pixel point in the initial ground surface coverage image corresponds to initial label information indicating the ground surface coverage object type; determining target label information corresponding to each pixel point in the earth surface sample image based on the initial label information corresponding to each pixel point in the initial earth surface covering image and the earth surface sample image; predicting a first probability that each pixel point in the surface sample image belongs to each of a plurality of preset surface covering categories based on the surface sample image and a first neural network; and adjusting the network parameter value in the first neural network based on the first probability corresponding to each pixel point and the target label information to obtain a target neural network.
In the embodiment of the disclosure, the initial label information indicating the earth surface covering object category at the pixel level is set, the more accurate target label information corresponding to each sample pixel point in the earth surface sample image is obtained by adjusting the initial label information, and then the first neural network is trained based on the earth surface sample image and the target label information, so that the trained target neural network can achieve higher accuracy when identifying the earth surface covering object.
In one possible embodiment, the determining, based on the initial tag information corresponding to each pixel point in the initial surface coverage image and the surface sample image, target tag information corresponding to each pixel point in the surface sample image includes: predicting a second probability that each pixel point in the surface sample image belongs to each surface covering category based on the surface sample image and a second neural network to be trained; adjusting network parameter values in the second neural network based on the second probability corresponding to each pixel point and the initial label information to obtain an initial second neural network; and determining target label information corresponding to each pixel point in the surface sample image based on the surface sample image, the initial second neural network and the initial label information corresponding to each pixel point.
In the embodiment, the initial label information corresponding to each pixel point in the surface sample image is adjusted through the initial second neural network obtained through training in advance, so that the target label information can be quickly and accurately screened without manual screening.
In one possible embodiment, the determining, based on the surface sample image, the initial second neural network, and the initial tag information corresponding to each pixel point, target tag information corresponding to each pixel point in the surface sample image includes: inputting the earth surface sample image into a current second neural network to obtain a second probability that each pixel point in the earth surface sample image predicted at the current time belongs to each earth surface covering category; the current second neural network is the initial second neural network or a neural network obtained by adjusting the initial second neural network at least once; determining current adjusted label information corresponding to each pixel point in the earth surface sample image based on a second probability that each pixel point in the current predicted earth surface sample image belongs to each earth surface covering category; adjusting the network parameter values in the current second neural network based on the second probability corresponding to each pixel point predicted at the current time, the initial label information corresponding to each pixel point and the label information corresponding to each pixel point after the last adjustment to obtain an adjusted current second neural network; and under the condition that a preset cut-off condition is not met, returning to the step of inputting the earth surface sample image into the current second neural network, and taking the label information of each pixel point after the current adjustment as the target label information corresponding to the pixel point after the preset cut-off condition is met.
In the embodiment, on one hand, the label information corresponding to each pixel point in the surface sample image is dynamically updated, on the other hand, the current second neural network is continuously optimized and adjusted through the dynamically updated label information, then, the label information corresponding to each pixel point is continuously and dynamically updated through the adjusted current second neural network, and the target label information with higher accuracy can be obtained through a mode of alternately updating the label information and the current second neural network.
In a possible embodiment, the preset cutoff condition comprises: at least one of the current adjustment times reaches a set time threshold value and the loss value corresponding to the current second neural network is smaller than the set threshold value, and target pixel points with the number exceeding the set number exist in the surface sample image; and the corresponding label information after the current adjustment of the target pixel point is the same as the corresponding label information after the last adjustment.
In the embodiment, for the convergence condition of the second neural network, on the basis of setting the convergence condition of the adjustment times and/or the network loss, the convergence condition that the label information after the current adjustment corresponding to the pixel point of the surface sample image tends to be stable is also added, so that the finally trained second neural network can achieve higher identification accuracy.
In a possible implementation manner, the determining, based on the second probability that each pixel point predicted at the current time belongs to each surface covering category, current adjusted tag information corresponding to each pixel point in the surface sample image includes: determining the confidence coefficient that each pixel point belongs to the target earth surface covering category based on the second probability that the pixel point respectively belongs to different earth surface covering categories aiming at each pixel point predicted at the current time; the target earth surface covering category is the earth surface covering category with the maximum corresponding probability; and determining the current adjusted label information corresponding to each pixel point in the earth surface sample image based on the confidence coefficient of each pixel point of the current prediction belonging to the category of the target earth surface covering and the confidence coefficient threshold corresponding to the earth surface sample image.
In the embodiment, the adjusted label information is determined by introducing the confidence coefficient that the pixel point belongs to the category of the target earth surface covering object, so that the accuracy of the finally determined target label information can be improved.
In a possible implementation manner, the determining, for each pixel point of the current prediction, a confidence that the pixel point belongs to the target earth surface covering category based on the second probabilities that the pixel point respectively belongs to different earth surface covering categories includes: acquiring the maximum probability and the second maximum probability in the second probability that the pixel point belongs to each earth surface covering category; and taking the ratio of the maximum probability to the second maximum probability as the confidence coefficient that the pixel point belongs to the category of the target earth surface covering object.
In one possible embodiment, the confidence threshold is determined according to the following steps: calculating the confidence coefficient mean value of each pixel point in the earth surface sample image belonging to the category of the target earth surface covering; taking the confidence coefficient mean value as the confidence coefficient threshold value when the confidence coefficient mean value is less than or equal to a set threshold value; and taking the set threshold as the confidence coefficient threshold when the confidence coefficient average value is larger than the set threshold.
In the above embodiment, different surface sample images have different confidence level means, so that the tag information corresponding to each surface sample image can be effectively updated when the confidence level means is less than or equal to the set threshold, and the robustness of the current second neural network is improved; in addition, when the confidence coefficient mean value is larger than the set threshold value, the threshold value is set to replace the confidence coefficient mean value as the confidence coefficient threshold value, so that the condition that the corresponding confidence coefficient is smaller than the confidence coefficient mean value but larger than the set threshold value is met, the label information of the pixel points is continuously updated, the label information of most of the pixel points can be continuously updated on the premise that the prediction result of the current second neural network is accurate, and the accuracy of the finally determined target label information can be improved.
In a possible implementation manner, the adjusting a network parameter value in the current second neural network based on the second probability corresponding to each pixel point predicted at the current time, the initial label information corresponding to each pixel point, and the label information corresponding to each pixel point after the last adjustment to obtain an adjusted current second neural network includes: determining a cross entropy loss value based on the second probability corresponding to each pixel point predicted at the current time and the label information corresponding to each pixel point after the last adjustment; determining a symmetric cross entropy loss value based on the second probability corresponding to each pixel point predicted at the current time and the initial label information corresponding to each pixel point; and determining a minimum entropy loss value based on the second probability corresponding to each pixel point of the current prediction; determining a total loss value corresponding to the current second neural network based on the cross entropy loss value, the symmetric cross entropy loss value and the minimized entropy loss value; and adjusting the network parameter value in the current second neural network based on the total loss value to obtain the adjusted current second neural network.
In the embodiment, the total loss value is determined by using the current prediction result of each pixel point, the last adjusted label information corresponding to each pixel point and the initial label information corresponding to each pixel point, and the current second neural network is adjusted by using the total loss value, so that the robustness of the adjusted current second neural network is improved.
In a third aspect, an embodiment of the present disclosure provides a surface covering identification apparatus, including:
the image acquisition module is used for acquiring a ground surface image; the category identification module is used for carrying out pixel point level classification identification on the earth surface image by using a target neural network to obtain category information of an earth surface covering object to which the pixel points in the earth surface image belong; the target neural network is obtained by training through a surface sample image and target label information corresponding to sample pixel points in the surface sample image, the target label information is used for identifying the surface covering object type to which the sample pixel points belong, and the target label information is obtained by adjusting initial label information corresponding to the sample pixel points in the surface sample image.
In a possible implementation manner, the apparatus for identifying a surface covering further includes an image generation module, and after the category video module obtains the category information of the surface covering to which the pixel points in the surface image belong, the image generation module is configured to:
determining the display color of the pixel point in the earth surface covering image corresponding to the earth surface image based on the category information corresponding to each pixel point in the earth surface image; and generating a ground surface covering image corresponding to the ground surface image based on the display color corresponding to each pixel point.
In a fourth aspect, an embodiment of the present disclosure provides a training apparatus for a neural network, including:
the system comprises a sample acquisition module, a data acquisition module and a data processing module, wherein the sample acquisition module is used for acquiring a ground surface sample image and an initial ground surface coverage image corresponding to the ground surface sample image, and each pixel point in the initial ground surface coverage image corresponds to initial label information indicating the type of a ground surface coverage object; the label determining module is used for determining target label information corresponding to each pixel point in the earth surface sample image based on the initial label information corresponding to each pixel point in the initial earth surface coverage image and the earth surface sample image; the probability prediction module is used for predicting a first probability that each pixel point in the earth surface sample image belongs to each earth surface covering category in multiple preset earth surface covering categories based on the earth surface sample image and the first neural network; and the parameter adjusting module is used for adjusting the network parameter value in the first neural network based on the first probability corresponding to each pixel point and the target label information to obtain a target neural network.
In one possible embodiment, the tag determining module, when configured to determine the target tag information corresponding to each pixel point in the surface sample image based on the initial tag information corresponding to each pixel point in the initial surface coverage image and the surface sample image, includes:
predicting a second probability that each pixel point in the surface sample image belongs to each surface covering category based on the surface sample image and a second neural network to be trained; adjusting network parameter values in the second neural network based on the second probability corresponding to each pixel point and the initial label information to obtain an initial second neural network; and determining target label information corresponding to each pixel point in the surface sample image based on the surface sample image, the initial second neural network and the initial label information corresponding to each pixel point.
In one possible embodiment, the label determining module, when configured to determine target label information corresponding to each pixel point in the surface sample image based on the surface sample image, the initial second neural network and the initial label information corresponding to each pixel point, includes:
inputting the earth surface sample image into a current second neural network to obtain a second probability that each pixel point in the earth surface sample image predicted at the current time belongs to each earth surface covering category; the current second neural network is the initial second neural network or a neural network obtained by adjusting the initial second neural network at least once; determining current adjusted label information corresponding to each pixel point in the earth surface sample image based on a second probability that each pixel point in the current predicted earth surface sample image belongs to each earth surface covering category; adjusting the network parameter values in the current second neural network based on the second probability corresponding to each pixel point predicted at the current time, the initial label information corresponding to each pixel point and the label information corresponding to each pixel point after the last adjustment to obtain an adjusted current second neural network; and under the condition that a preset cut-off condition is not met, returning to the step of inputting the earth surface sample image into the current second neural network, and taking the label information of each pixel point after the current adjustment as the target label information corresponding to the pixel point after the preset cut-off condition is met.
In a possible embodiment, the preset cutoff condition comprises:
at least one of the current adjustment times reaches a set time threshold value and the loss value corresponding to the current second neural network is smaller than the set threshold value, and target pixel points with the number exceeding the set number exist in the surface sample image; and the corresponding label information after the current adjustment of the target pixel point is the same as the corresponding label information after the last adjustment.
In a possible embodiment, the tag determining module, when configured to determine current adjusted tag information corresponding to each pixel point in the surface sample image based on the second probability that each pixel point predicted at the current time belongs to each surface covering category, includes:
determining the confidence coefficient that each pixel point belongs to the target earth surface covering category based on the second probability that the pixel point respectively belongs to different earth surface covering categories aiming at each pixel point predicted at the current time; the target earth surface covering category is the earth surface covering category with the maximum corresponding probability; and determining the current adjusted label information corresponding to each pixel point in the earth surface sample image based on the confidence coefficient of each pixel point of the current prediction belonging to the category of the target earth surface covering and the confidence coefficient threshold corresponding to the earth surface sample image.
In a possible embodiment, the tag determining module, when configured to determine, for each pixel point of the current prediction, a confidence that the pixel point belongs to the target earth surface covering category based on the second probabilities that the pixel point respectively belongs to different earth surface covering categories, includes:
acquiring the maximum probability and the second maximum probability in the second probability that the pixel point belongs to each earth surface covering category; and taking the ratio of the maximum probability to the second maximum probability as the confidence coefficient that the pixel point belongs to the category of the target earth surface covering object.
In one possible embodiment, the tag determination module is configured to determine the confidence threshold according to the following steps:
calculating the confidence coefficient mean value of each pixel point in the earth surface sample image belonging to the category of the target earth surface covering; taking the confidence coefficient mean value as the confidence coefficient threshold value when the confidence coefficient mean value is less than or equal to a set threshold value; and taking the set threshold as the confidence coefficient threshold when the confidence coefficient average value is larger than the set threshold.
In a possible implementation manner, the tag determining module, when configured to adjust a network parameter value in the current second neural network based on the second probability corresponding to each pixel point predicted at the current time, initial tag information corresponding to each pixel point, and last adjusted tag information corresponding to each pixel point, to obtain an adjusted current second neural network, includes:
determining a cross entropy loss value based on the second probability corresponding to each pixel point predicted at the current time and the label information corresponding to each pixel point after the last adjustment; determining a symmetric cross entropy loss value based on the second probability corresponding to each pixel point predicted at the current time and the initial label information corresponding to each pixel point; and determining a minimum entropy loss value based on the second probability corresponding to each pixel point of the current prediction; determining a total loss value corresponding to the current second neural network based on the cross entropy loss value, the symmetric cross entropy loss value and the minimized entropy loss value; and adjusting the network parameter value in the current second neural network based on the total loss value to obtain the adjusted current second neural network.
In a fifth aspect, an embodiment of the present disclosure provides an electronic device, including: a processor, a memory and a bus, the memory storing machine-readable instructions executable by the processor, the processor and the memory communicating over the bus when the electronic device is operating, the machine-readable instructions when executed by the processor performing the steps of the method of the first or second aspect.
In a sixth aspect, the disclosed embodiments provide a computer-readable storage medium having a computer program stored thereon, where the computer program is executed by a processor to perform the steps of the method according to the first or second aspect.
In order to make the aforementioned objects, features and advantages of the present disclosure more comprehensible, preferred embodiments accompanied with figures are described in detail below.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the drawings required for use in the embodiments will be briefly described below, and the drawings herein incorporated in and forming a part of the specification illustrate embodiments consistent with the present disclosure and, together with the description, serve to explain the technical solutions of the present disclosure. It is appreciated that the following drawings depict only certain embodiments of the disclosure and are therefore not to be considered limiting of its scope, for those skilled in the art will be able to derive additional related drawings therefrom without the benefit of the inventive faculty.
FIG. 1 illustrates a schematic diagram of a terrain surface image and a corresponding terrain surface overlay image provided by an embodiment of the present disclosure;
FIG. 2 illustrates a flow chart of a method of identifying a surface covering provided by an embodiment of the present disclosure;
FIG. 3 illustrates a flowchart of generating a corresponding earth surface coverage image of an earth surface image according to an embodiment of the present disclosure;
FIG. 4 is a flow chart illustrating a method for training a neural network provided by an embodiment of the present disclosure;
fig. 5 is a flowchart illustrating a method for determining target tag information according to an embodiment of the present disclosure;
fig. 6 is a flowchart illustrating a specific determination method of target tag information according to an embodiment of the present disclosure;
FIG. 7 is a flowchart illustrating a method for determining current adjusted tag information corresponding to each pixel point in an image of a surface sample according to an embodiment of the present disclosure;
FIG. 8 is a flow chart illustrating a method for determining the confidence level that a pixel belongs to a category of a target surface covering according to an embodiment of the present disclosure;
FIG. 9 illustrates a flow chart for determining a confidence threshold provided by an embodiment of the present disclosure;
FIG. 10 is a schematic flow chart illustrating a method for determining target tag information according to an embodiment of the present disclosure;
FIG. 11 is a schematic structural diagram illustrating a device for identifying a surface covering provided by an embodiment of the present disclosure;
fig. 12 is a schematic structural diagram illustrating a training apparatus for a neural network provided in an embodiment of the present disclosure;
fig. 13 is a schematic structural diagram of an electronic device provided in an embodiment of the present disclosure;
fig. 14 shows a schematic structural diagram of another electronic device provided in the embodiment of the present disclosure.
Detailed Description
In order to make the objects, technical solutions and advantages of the embodiments of the present disclosure more clear, the technical solutions of the embodiments of the present disclosure will be described clearly and completely with reference to the drawings in the embodiments of the present disclosure, and it is obvious that the described embodiments are only a part of the embodiments of the present disclosure, not all of the embodiments. The components of the embodiments of the present disclosure, generally described and illustrated in the figures herein, can be arranged and designed in a wide variety of different configurations. Thus, the following detailed description of the embodiments of the present disclosure, presented in the figures, is not intended to limit the scope of the claimed disclosure, but is merely representative of selected embodiments of the disclosure. All other embodiments, which can be derived by a person skilled in the art from the embodiments of the disclosure without making creative efforts, shall fall within the protection scope of the disclosure.
It should be noted that: like reference numbers and letters refer to like items in the following figures, and thus, once an item is defined in one figure, it need not be further defined and explained in subsequent figures.
As shown in fig. 1 (a), in order to set a high-resolution ground surface image corresponding to a geographic area, as shown in fig. 1 (b), after image processing is performed on a low-resolution ground surface coverage image corresponding to the set geographic area, an up-sampling process is performed on the low-resolution ground surface coverage image through, for example, a nearest neighbor difference algorithm to obtain a corresponding high-resolution ground surface coverage image, then a neural network for identifying a ground surface coverage type is trained based on the high-resolution ground surface image and the corresponding high-resolution ground surface coverage image, and then a ground surface coverage type corresponding to a high-resolution ground surface image corresponding to another geographic area is predicted based on the neural network, so that the high-resolution ground surface coverage image can be obtained based on the ground surface coverage type.
However, the low-resolution ground surface coverage image is automatically or semi-automatically generated by an algorithm, a pixel point with a ground surface coverage category recognized incorrectly exists, a label corresponding to the pixel with the ground surface coverage category recognized incorrectly is a noise label, and in addition, when the low-resolution ground surface coverage image is subjected to image up-sampling processing, a part of the noise label is introduced, so that a large number of noise labels may exist in the obtained high-resolution ground surface coverage image, and thus when the neural network is trained based on the high-resolution ground surface coverage image containing the noise label, the recognition accuracy of the neural network when the neural network is applied to the ground surface coverage recognition process is low.
In view of the above, the embodiment of the present disclosure provides a method for identifying a surface covering, in which a pre-trained target neural network is used to perform classification and identification of pixel levels on a surface image, and in addition, in a training phase, the target neural network is obtained by training based on target label information corresponding to a sample pixel point in a surface sample image, and the target label information obtained by adjusting initial label information is more accurate, so that when the classification information of the surface covering to which each pixel point in the surface image belongs is predicted based on the target neural network, the identification accuracy of the surface covering can be improved.
At least one embodiment of the present disclosure provides a method of surface covering identification, which may be performed by a computer device having certain computing capabilities, the computer device comprising, for example: terminal equipment or servers or other processing devices. In some possible implementations, the surface covering identification method may be implemented by a processor invoking computer readable instructions stored in a memory.
Referring to fig. 2, a flowchart of a method for identifying a surface covering according to an embodiment of the present disclosure is shown, where the method includes the following steps S201 to S202:
s201, acquiring a ground surface image.
The earth surface image may be, for example, an earth surface image of a set geographical region in a high-resolution satellite imagery.
S202, carrying out classification and identification of pixel point levels on the earth surface image by using the target neural network to obtain the category information of the earth surface covering object to which the pixel points in the earth surface image belong. The target neural network is obtained by training through the earth surface sample image and target label information corresponding to sample pixel points in the earth surface sample image, the target label information is used for identifying the earth surface covering object type to which the sample pixel points belong, and the target label information is obtained by adjusting initial label information corresponding to the sample pixel points in the earth surface sample image.
The training process for the target neural network and the determination process for the target label information will be described in detail below.
Illustratively, the earth surface image may be input into a target neural network, the target neural network performs feature extraction processing on the earth surface image to obtain feature information of each pixel point in the earth surface image, illustratively, the feature information of each pixel point may include feature information of texture features, spectral features and the like of the pixel point, and the target neural network may determine category information of an earth surface covering to which the pixel point belongs based on the feature information of each pixel point.
Illustratively, the category information of the surface covering may include, but is not limited to, one of a land category, a building category, a grass category, a water body category, and the like. The land category, the building category, the grassland category, and the water body category may be further subdivided into sub-categories, which is not limited in the present disclosure.
In summary, the embodiment of the present disclosure may perform classification and identification at a pixel level on the surface image by using a pre-trained target neural network, and in addition, the target neural network is trained based on target label information corresponding to a sample pixel point in the surface sample image in a training stage, so as to improve the identification accuracy of the surface covering.
Further, as shown in fig. 3, after obtaining the category information of the earth surface covering to which the pixel points in the earth surface image belong, the method for identifying the earth surface covering provided by the embodiment of the present disclosure further includes:
s203, determining the display color of the pixel point in the earth surface covering image corresponding to the earth surface image based on the category information corresponding to each pixel point in the earth surface image;
and S204, generating a ground surface covering image corresponding to the ground surface image based on the display color corresponding to each pixel point.
The image of the ground surface covering included in the ground surface can be visually represented by different colors, for example, green represents grassland, yellow represents land, blue represents water area, brown represents buildings, and the like.
Illustratively, a mapping relationship between the category information and the display color may be established in advance, and the display color to be presented of each pixel point in the earth surface coverage image is determined based on the category information corresponding to the pixel point, so as to obtain the earth surface coverage image corresponding to the earth surface image.
For example, for a surface image including 100 × 100 pixels, if it is determined that the category information of the pixels in the first row and the first column is a grassland category, and in the mapping relationship between the pre-established category information and the display colors, the display colors corresponding to the grassland category are green, the pixels in the first row and the first column in the surface overlay image are green, and according to the same manner, the display colors of each pixel in the surface image in the surface overlay image can be determined, and further the surface overlay image corresponding to the surface image can be generated.
In the embodiment of the disclosure, the display colors of the pixel points are respectively determined according to the category information corresponding to the pixel points, so that the areas where the earth surface coverings of different categories are located can be clearly identified in the earth surface chart.
As shown in fig. 4, a training method for a neural network provided in the embodiment of the present disclosure is used to obtain a target neural network mentioned in the above embodiment, and specifically includes the following steps S401 to S404:
s401, acquiring a surface sample image and an initial surface covering image corresponding to the surface sample image, wherein each pixel point in the initial surface covering image corresponds to initial label information indicating the type of the surface covering.
For example, the low-resolution surface coverage image corresponding to each surface sample image may be obtained in advance, and then the low-resolution surface coverage image corresponding to each surface sample image is subjected to image processing to obtain the initial surface coverage image corresponding to the surface sample image.
For example, for a low-resolution earth surface coverage image corresponding to any one earth surface sample image, the low-resolution earth surface coverage image may be processed according to a nearest neighbor difference algorithm, for example, after a new element is inserted between pixel points in the low-resolution earth surface coverage image according to the nearest neighbor difference algorithm, an initial earth surface coverage image corresponding to any one earth surface sample image may be obtained, and the obtained initial earth surface coverage image is the same as the number of pixel points in any one earth surface sample image.
For example, the initial label information corresponding to the pixel points in the initial ground surface coverage image may be used to identify the ground surface coverage categories, for example, the initial label information may be represented in a one-hot vector form, where the dimensions of the one-hot vector are the same as the number of the ground surface coverage categories, for example, the ground surface coverage categories include four types in total, which are respectively a land category, a building category, a grass category, and a water body category, where the one-hot vector is a four-dimensional vector, each dimension corresponds to one ground surface coverage category, specifically, the correspondence between the dimension in the one-hot vector and the ground surface coverage category may be preset, for example, in the one-hot vector, the first dimension corresponds to the land category, the second dimension corresponds to the building category, the third dimension corresponds to the grass category, and the fourth dimension corresponds to the water body category, when the one-hot vector is used for representing the initial label information, determining the numerical value of each dimension in the one-hot vector according to the earth surface covering type identified by the initial label information, wherein if the earth surface covering type identified by the initial label information is the land type, the one-hot vector is [1000 ]; if the ground cover type identified by the initial tag information is a building type, the one-hot vector is [1000 ].
Exemplarily, if a first dimension in a one-hot vector corresponds to a land category, a second dimension corresponds to a building category, a third dimension corresponds to a grassland category, and a fourth dimension water category, if initial label information of any pixel point in an initial ground surface coverage image is [1000 ], the ground surface coverage category of the any pixel point is the land category.
S402, determining target label information corresponding to each pixel point in the surface sample image based on the initial label information corresponding to each pixel point in the initial surface covering image and the surface sample image.
The initial label information corresponding to the pixel points can be called as noise labels, and here, the noise labels are continuously adjusted to obtain target label information corresponding to each pixel point in the surface sample image, and how to adjust to obtain the target label information corresponding to each pixel point will be described in detail later.
S403, predicting a first probability that each pixel point in the surface sample image belongs to each surface covering category in multiple preset surface covering categories based on the surface sample image and the first neural network.
For example, after the surface sample image is input into the first neural network, the first neural network may perform image processing on the surface sample image, extract feature information of each pixel point in the surface sample image, and predict a first probability that the pixel point belongs to each of the plurality of preset surface cover categories based on the feature information of the pixel point.
Exemplarily, the first probability that each pixel point belongs to each of the preset surface covering categories can represent the possibility that the pixel point belongs to each of the preset surface covering categories, the first probability corresponding to each pixel point can form a probability vector, the number of dimensions contained in the probability vector, the surface covering category corresponding to each dimension is consistent with the one-hot vector mentioned above, and the probability value of each dimension can represent the probability that the pixel point belongs to the surface covering category represented by the dimension.
For example, after the surface sample image is input into the first neural network, if the first probability corresponding to any pixel in the surface sample image is [ 0.60.20.10.1 ], the first probability indicating that any pixel belongs to the land category is 0.6, the first probability indicating that any pixel belongs to the building category is 0.2, the first probability indicating that any pixel belongs to the grass category is 0.1, and the first probability indicating that any pixel belongs to the water category is 0.1.
S404, adjusting network parameter values in the first neural network based on the first probability corresponding to each pixel point and the target label information to obtain the target neural network.
For example, when the target label information is predicted for the pixel points of the surface sample image, a loss function between the obtained first probability corresponding to each pixel point and the target label information corresponding to the pixel point may be determined based on the first probability and the target label information, and then the network parameter value in the first neural network is adjusted based on the loss function until the loss value corresponding to the first neural network is smaller than a set loss threshold, or the training times reach a set time threshold, so as to obtain the above-mentioned target neural network.
In the embodiment of the disclosure, the initial label information indicating the earth surface covering object category at the pixel level is set, the more accurate target label information corresponding to each sample pixel point in the earth surface sample image is obtained by adjusting the initial label information, and then the first neural network is trained based on the earth surface sample image and the target label information, so that the trained target neural network can achieve higher accuracy when identifying the earth surface covering object.
How to obtain the target label information corresponding to each pixel point in the surface sample image is described in detail below.
Specifically, when the target label information corresponding to each pixel point in the surface sample image is determined based on the initial label information corresponding to each pixel point in the initial surface coverage image and the surface sample image, as shown in fig. 5, the following S501 to S503 are included:
s501, predicting a second probability that each pixel point in the surface sample image belongs to each surface covering object type based on the surface sample image and a second neural network to be trained.
For example, after the surface sample image is input into the second neural network to be trained, the second neural network may perform image processing on the surface sample image, extract feature information of each pixel point in the surface sample image, and predict a second probability that the pixel point belongs to each of the plurality of preset surface cover categories based on the feature information of the pixel point.
For example, the second neural network to be trained and the first neural network are both used for identifying the category of the surface covering, and the surface sample image is input into the second neural network to be trained, so that the second probability that each pixel point in the surface sample image belongs to each category of the surface covering can be obtained.
Illustratively, the second probabilities corresponding to the pixel points are similar to the first probabilities corresponding to the pixel points, and each can be represented by a probability vector, and the number of dimensions included in the probability vector formed by the second probabilities, and the ground surface covering category represented by each dimension are the same as the probability vector formed by the first probabilities above.
S502, adjusting network parameter values in the second neural network based on the second probability and the initial label information corresponding to each pixel point to obtain an initial second neural network.
Considering that the initial label information includes a noise label, in order to avoid overfitting of the obtained initial second neural network as much as possible, when the network parameter values in the second neural network are adjusted based on the second probability corresponding to each pixel point and the initial label information, the training may be stopped before the loss function corresponding to the second neural network is not completely converged.
For example, if the loss function is completely converged, the corresponding training frequency should reach the set frequency threshold N times, the training may be ended in advance when the training frequency is less than N times, or, if the loss function is completely converged, the training may be ended when the loss value corresponding to the second neural network is less than the set loss threshold, the training may be ended in advance when the loss value is not less than the preset loss threshold, so as to obtain an initial second neural network that is not completely converged.
S503, determining target label information corresponding to each pixel point in the surface sample image based on the surface sample image, the initial second neural network and the initial label information corresponding to each pixel point.
In the embodiment of the disclosure, the initial label information corresponding to each pixel point in the surface sample image is adjusted through the initial second neural network obtained through training in advance, so that the target label information can be quickly and accurately screened without manual screening.
For example, the specific scheme for adjusting the target tag information corresponding to each pixel point in the surface sample image, as shown in fig. 6, may include the following steps S601 to S604:
s601, inputting the earth surface sample image into a current second neural network to obtain a second probability that each pixel point in the current predicted earth surface sample image belongs to each earth surface covering category.
The current second neural network is the initial second neural network or the neural network obtained by adjusting the initial second neural network at least once.
The second probability corresponding to each pixel point may be obtained by inputting the surface sample image into the initial second neural network obtained by the training in S402 for processing, or by inputting the surface sample image into the neural network obtained by at least one adjustment of the initial second neural network for processing, and how to adjust the initial second neural network is described in detail below.
The current second neural network can also perform image processing on the surface sample image, extract the characteristic information of each pixel point in the surface sample image, and predict the second probability that the pixel point belongs to each surface covering category based on the characteristic information of the pixel point.
S602, determining current adjusted label information corresponding to each pixel point in the earth surface sample image based on the second probability that each pixel point in the current predicted earth surface sample image belongs to each earth surface covering object category.
For example, the sum of the second probabilities that each pixel point in the earth surface sample image predicted each time belongs to each earth surface covering category is 1, and for example, the current adjusted label information corresponding to each pixel point may be determined based on the earth surface covering category corresponding to the maximum probability of the second probabilities obtained by each pixel point in the current prediction.
Specifically, for any pixel point, if the last adjusted label information corresponding to the pixel point indicates that the earth surface covering category of the pixel point is the land category, and the second probability that the pixel point obtained by the current prediction belongs to the grassland category is the highest, the current adjusted label information corresponding to the pixel point may become the grassland category, specifically, whether it is determined whether the current adjusted label information corresponding to the pixel point is changed into the grassland category is related to the confidence of the prediction result obtained by predicting the pixel point, the confidence is higher, the current adjusted label information corresponding to the pixel point may become the grassland category, if the confidence is lower, the current adjusted label information corresponding to the pixel point may still be the land category, specifically, how to determine based on the confidence is, see the description below for details.
Particularly, if the current prediction is that the earth surface covering category of a pixel point in the earth surface sample image is predicted based on the initial second neural network for the first time, the label information after the last adjustment is the initial label information of the pixel point
S503, adjusting the network parameter values in the current second neural network based on the second probability corresponding to each pixel point predicted at the current time, the initial label information corresponding to each pixel point and the label information corresponding to each pixel point after the last adjustment, so as to obtain the adjusted current second neural network.
Illustratively, when the network parameter value in the current second neural network is adjusted, the network parameter value in the network is adjusted as follows:
(1) based on the second probability corresponding to each pixel point predicted at the current time and the initial label information of the pixel point, the determined symmetric cross entropy loss function can avoid that the current second neural network is over-fitted to the updated label aiming at the label information prediction result of the pixel point, so that the deviation with the initial label information is larger, and the problem that the prediction accuracy of the current second neural network is reduced is caused.
(2) And determining a minimum entropy loss function based on a second probability corresponding to each pixel point of the current prediction, and adjusting the network parameter value of the current second neural network by adjusting the loss value corresponding to the minimum entropy loss function so as to improve the robustness of the current second neural network.
(3) And determining a cross entropy loss function representing the difference between the current label information and the last adjusted label information of the pixel point based on the second probability corresponding to each pixel point of the current prediction and the last adjusted label information corresponding to the pixel point, and reducing the error between the current prediction result and the last adjusted label information by adjusting the loss value corresponding to the cross entropy loss function so as to improve the robustness of the current second neural network.
And determining a total loss function corresponding to the current second neural network based on the symmetric cross entropy loss function, the minimized entropy loss function and the cross entropy loss function, and then adjusting each network parameter value in the current second neural network based on the loss value corresponding to the total loss function to obtain the adjusted current second neural network.
And S604, under the condition that the preset cut-off condition is not met, returning to the step of inputting the earth surface sample image into the current second neural network, and taking the label information of each pixel point after the current adjustment as the target label information corresponding to the pixel point after the preset cut-off condition is met.
Here, when it is determined that the preset cutoff condition is not satisfied, the step of inputting the surface sample image into the current second neural network is returned, and the current second neural network at this time is the adjusted current second neural network obtained in S603, and may also be understood as a neural network obtained by performing at least one adjustment on the initial second neural network.
In the embodiment of the disclosure, on one hand, the label information corresponding to each pixel point in the surface sample image is dynamically updated, on the other hand, the current second neural network is continuously optimized and adjusted through the dynamically updated label information, then, the label information corresponding to each pixel point is continuously and dynamically updated through the adjusted current second neural network, and the target label information with higher accuracy can be obtained through a mode of alternately updating the label information and the current second neural network.
Illustratively, the preset cutoff condition includes the following condition (1) and condition (2):
(1) the current adjusting times reach at least one of a set time threshold value and a loss value corresponding to the current second neural network is smaller than the set threshold value;
(2) target pixel points with more than a set number exist in the earth surface sample image; and the corresponding label information after the current adjustment of the target pixel point is the same as the corresponding label information after the last adjustment.
Here, when the condition (1) and the condition (2) are satisfied at the same time, the preset cutoff condition is considered to be satisfied, whereas the preset cutoff condition is considered to be not satisfied.
In the condition (1), when the current adjustment number reaches a set number threshold, or a loss value corresponding to the current second neural network (i.e., a loss value corresponding to the aforementioned total loss function) is smaller than the set threshold, it indicates that the current second neural network reaches convergence; in the condition (2), after two adjacent adjustments, the label information of target pixels with more than the set number is the same, which indicates that the label information of each pixel after the current adjustment is relatively stable and accurate, and at this time, the label information of each pixel after the current adjustment can be used as the target label information corresponding to the pixel.
In the embodiment, for the convergence condition of the second neural network, on the basis of setting the convergence condition of the adjustment times and/or the network loss, the convergence condition that the label information after the current adjustment corresponding to the pixel point of the surface sample image tends to be stable is also added, so that the finally trained second neural network can achieve higher identification accuracy.
Specifically, when determining the currently adjusted label information corresponding to each pixel point in the surface sample image based on the second probability that each pixel point predicted at the current time belongs to each surface covering category, as shown in fig. 7, the method may include the following steps S701 to S702:
s701, aiming at each pixel point predicted at the current time, determining the confidence coefficient that the pixel point belongs to the category of the target earth surface covering object based on the second probability that the pixel point respectively belongs to different categories of the earth surface covering object; the target earth surface covering category is the earth surface covering category with the maximum corresponding probability;
s702, determining current adjusted label information corresponding to each pixel point in the earth surface sample image based on the confidence coefficient that each pixel point predicted at the current time belongs to the target earth surface covering object category and the confidence coefficient threshold corresponding to the earth surface sample image.
In view of the above, the currently adjusted tag information corresponding to each pixel point may be determined based on the earth surface covering category corresponding to the maximum probability in the second probability obtained by current prediction for each pixel point, and further, confidence is introduced to determine the reliability of the pixel point belonging to the earth surface covering category corresponding to the maximum probability.
In the embodiment of the disclosure, the adjusted label information is determined by introducing the confidence that the pixel point belongs to the category of the target earth surface covering object, so that the accuracy of the finally determined target label information can be improved.
Specifically, when determining, for each pixel point of the current prediction, the confidence that the pixel point belongs to the target earth surface cover category based on the second probabilities that the pixel point respectively belongs to different earth surface cover categories, as shown in fig. 8, the following steps S801 to S802 may be included:
s801, acquiring the maximum probability and the second maximum probability in the second probability that the pixel point belongs to each earth surface covering category;
s802, taking the ratio of the maximum probability to the second maximum probability as the confidence coefficient that the pixel point belongs to the category of the target earth surface covering object.
For example, if the preset earth surface covering category includes four categories, where the second probability corresponding to the pixel point is four, and the maximum probability and the second probability are selected from the four second probabilities, for example, a probability vector formed by the second probability corresponding to the pixel point is [ 0.60.20.10.1 ], a ratio 3 between the maximum probability 0.6 and the second probability 0.2 is used as the confidence that the pixel point belongs to the target earth surface covering category.
For the above S702, when it is determined that the confidence that the current predicted pixel point belongs to the target earth surface cover category reaches the confidence threshold, the target earth surface cover category is used as the current adjusted tag information corresponding to the pixel point; on the contrary, under the condition that the confidence coefficient that the pixel point predicted at the current time belongs to the target earth surface covering category does not reach the confidence coefficient threshold value, the earth surface covering category of the pixel point still adopts the earth surface covering category identified by the label information adjusted at the last time, and the label information of the pixel point is not updated at the current time.
Exemplarily, for any pixel point, if the label information corresponding to the pixel point after the last adjustment is a grassland category, if the confidence coefficient that any pixel point predicted at the current time belongs to a land category is 9, the confidence coefficient threshold is 7, and the confidence coefficient is greater than the confidence coefficient threshold, determining that the label information corresponding to the pixel point after the current adjustment is a one-hot vector representing the land category; and if the confidence coefficient that any pixel point of the current prediction belongs to the land category is 6 and the confidence coefficient is smaller than the confidence coefficient threshold value, determining that the label information corresponding to any pixel point is still the grassland category.
Specifically, the confidence threshold may be determined according to the following steps, as shown in fig. 9, which may include S901 to S903:
s901, calculating a confidence coefficient mean value of each pixel point in the earth surface sample image belonging to the category of the target earth surface covering;
s902, taking the confidence coefficient mean value as a confidence coefficient threshold value under the condition that the confidence coefficient mean value is less than or equal to the set threshold value;
and S903, taking the set threshold as the confidence threshold when the confidence mean value is larger than the set threshold.
For example, after determining the confidence corresponding to each pixel point in the surface sample image, a confidence map corresponding to the surface sample image may be generated, where each pixel point in the confidence map is represented by the confidence corresponding to the pixel point, and then the confidence mean corresponding to the confidence map may be further determined, so that it can be seen that the confidence mean corresponding to each surface sample image is related to the confidence of the pixel points included in the surface sample image, and each surface sample image corresponds to a unique confidence mean.
Therefore, when the confidence coefficient mean value is smaller than or equal to the set threshold value, the label information of the pixel points with the confidence coefficient reaching the confidence coefficient mean value is updated for each earth surface sample image, and the label information of the pixel points with the confidence coefficient not reaching the confidence coefficient mean value is not updated, so that the label information of the pixel points can be effectively updated based on the confidence coefficient mean value corresponding to each confidence coefficient image, the robustness of the current second neural network can be improved, and accurate target label information can be finally obtained.
In addition, when the confidence coefficient mean value is larger than the set threshold value, the threshold value is set to replace the confidence coefficient mean value as the confidence coefficient threshold value, so that the condition that the corresponding confidence coefficient is smaller than the confidence coefficient mean value but larger than the set threshold value is met, the label information of the pixel points is continuously updated, the label information of most of the pixel points can be continuously updated on the premise that the prediction result of the current second neural network is accurate, and the accuracy of the finally determined target label information can be improved.
After obtaining the label information after the current adjustment corresponding to each pixel point in the surface sample image, the following describes in detail with respect to the above S603, how to specifically adjust the network parameter value in the current second neural network:
introducing a symmetric cross entropy loss function representing the difference between the prediction result of the current label information of the pixel point and the initial label information, wherein the value corresponding to the symmetric cross entropy loss function passes through a symmetric cross entropy loss value LsceExpressing, the symmetric cross entropy loss value is composed of two parts, one part is a cross entropy loss value L between the prediction result of the current label information of the characterization pixel point and the initial label informationceThe other part is reverse intersection between the prediction result of the current label information of the characterization pixel point and the initial label informationFork entropy loss value LrceSpecifically, the symmetric cross-entropy loss value, the cross-entropy loss value, and the inverse cross-entropy loss value may be represented by the following formulas (1), (2), and (3):
Lsce=Lce+ωLrce(1);
the method comprises the steps of inputting pixel points x, wherein x represents characteristic information of the input pixel points x, the characteristic information can comprise texture information, spectrum information and the like of the input pixel points x and can reflect the characteristic information of the pixel points, y ∈ { 1., K } represents prediction label categories corresponding to the pixel points, K represents the number of the label categories, K represents variables for indicating ground surface covering object categories, K ∈ { 1., K }, and omega represents an inverse cross entropy loss value LrceIn determining symmetric cross entropy loss value LsceA weight of time; p (k | x) represents a probability value that the earth surface covering type of the pixel point x obtained after the label information of the pixel point x is predicted belongs to the earth surface covering type indicated by k, q (k | x) represents a probability value that the earth surface covering type of the pixel point x belongs to the earth surface covering type indicated by k in a one-hot vector corresponding to the initial label information of the pixel point x (q (k | x) is equal to 1 or 0), and exemplarily, if the one-hot vector corresponding to the initial label information of the pixel point x is [1000 []If k is 2, the probability value indicating that the earth surface covering category of the pixel point x belongs to the second earth surface covering category is obtained, and the probability value is 0 based on the one-hot vector.
For the above-mentioned minimized entropy loss function, a minimized entropy loss value corresponding to the minimized entropy loss function is introduced, and the minimized entropy loss value is determined based on the second probability corresponding to each pixel point of the current prediction, which can be specifically represented by the following formula (4):
wherein L iseThe minimum loss value corresponding to the minimum loss function is represented.
For the cross entropy loss function which represents the difference between the current label information of the pixel point and the label information adjusted last time, a loss value corresponding to the cross entropy loss function is called a cross entropy loss value, and can be specifically determined by the following formula (5):
wherein, L'ceRepresenting a cross entropy loss value corresponding to the cross entropy loss function;after the label prediction is carried out on the pixel point x, in the one-hot vector corresponding to the current label information of the pixel point x, the probability value that the earth surface covering category of the pixel point x belongs to the earth surface covering category indicated by the k is shownEqual to 1 or 0). Exemplarily, if the one-hot vector corresponding to the current tag information of the pixel point x is [1000 ]]If k is 1, the probability value indicating that the earth surface covering category of the pixel point x belongs to the first earth surface covering category is obtained, and the probability value is 0 based on the one-hot vector.
Further, based on the cross entropy loss value, the symmetrical cross entropy loss value and the minimized entropy loss value, a total loss value L corresponding to the current second neural network is determinedtotalSpecifically, it can be determined by the following formula (6):
Ltotal=L′ce+αLsce+βLe(6);
where α and β are weighting parameters, which may be preset, for example, α may be 0.5 and β may be 1, and then the total loss value may be determined by adjusting each network parameter value included in the total loss function.
Further, after the total loss value is obtained, the network parameter value in the current second neural network may be adjusted based on the total loss value, so as to obtain an adjusted current second neural network.
In the embodiment of the disclosure, the total loss value is determined by using the current prediction result of each pixel point, the last adjusted label information corresponding to each pixel point, and the initial label information corresponding to each pixel point, and the current second neural network is adjusted by using the total loss value, so that the robustness of the adjusted current second neural network is improved.
A specific embodiment is given below in conjunction with fig. 10, and a summary description is given for the above target pixel labels for determining each pixel point in the surface sample image:
the CNN neural network in fig. 10 includes the initial second neural network mentioned above, or the neural network after at least one adjustment to the initial second neural network, where the CNN neural network includes a confidence updating module, the confidence updating module is configured to dynamically update label information corresponding to each pixel point in the surface sample image, after the surface sample image is input into the CNN neural network, a second probability that each pixel point belongs to each surface covering category can be obtained, and can be represented by the prediction result p in fig. 10, then a symmetric cross entropy loss function is obtained based on the prediction result and the initial label information (represented by the original noisy label in fig. 10), a minimized entropy loss function is obtained based on the prediction result, and a symmetric cross entropy loss function is obtained based on the prediction result and the label information after the last adjustment (represented by the dynamically updated label in fig. 10, during initialization, the dynamically updated label is an original label with noise), cross entropy loss functions are obtained, the CNN neural network is adjusted based on the loss functions until the CNN network converges, and target label information corresponding to each pixel point in the surface sample image is obtained under the condition that label information of target pixel points with more than a set number exists in the pixel points in the obtained surface sample image after the CNN neural network is adjusted for two adjacent times.
It will be understood by those skilled in the art that the order of description of the steps in the above method is not meant to be a strict order of execution and to constitute any limitation on the implementation, and the specific order of execution of the steps should be determined by their function and possible inherent logic.
Based on the same technical concept, the embodiment of the present disclosure further provides a surface covering identification device corresponding to the surface covering identification method, and as the principle of solving the problem of the device in the embodiment of the present disclosure is similar to that of the surface covering identification method in the embodiment of the present disclosure, the implementation of the device may refer to the implementation of the method, and repeated details are not repeated.
Referring to fig. 11, a schematic diagram of a ground surface covering recognition apparatus 1100 provided in an embodiment of the present disclosure is shown, where the ground surface covering recognition apparatus 1100 includes: an image acquisition module 1101 and a category identification module 1102.
The image acquisition module 1101 is used for acquiring a ground surface image;
the category identification module 1102 is configured to perform classification identification of pixel levels on the earth surface image by using a target neural network to obtain category information of an earth surface covering to which the pixels in the earth surface image belong;
the target neural network is obtained by training through the earth surface sample image and target label information corresponding to sample pixel points in the earth surface sample image, the target label information is used for identifying the earth surface covering object type to which the sample pixel points belong, and the target label information is obtained by adjusting initial label information corresponding to the sample pixel points in the earth surface sample image.
In a possible embodiment, the ground cover identifying apparatus 1100 further includes an image generating module 1103, and after the category video module 1102 obtains the category information of the ground cover to which the pixel points in the ground image belong, the image generating module 1103 is configured to:
determining the corresponding display color of the pixel point in the earth surface covering image corresponding to the earth surface image based on the category information corresponding to each pixel point in the earth surface image;
and generating a ground surface covering image corresponding to the ground surface image based on the display color corresponding to each pixel point.
Based on the same technical concept, the embodiment of the present disclosure further provides a training apparatus for a neural network corresponding to the training method for the neural network, and since the principle of the apparatus in the embodiment of the present disclosure for solving the problem is similar to the training method in the embodiment of the present disclosure, the implementation of the apparatus may refer to the implementation of the method, and repeated details are not repeated.
Referring to fig. 12, a schematic diagram of a training apparatus 1200 of a neural network provided in an embodiment of the present disclosure is shown, where the training apparatus 1200 of the neural network includes: a sample acquisition module 1201, a label determination module 1202, a probability prediction module 1203, and a parameter adjustment module 1204.
The system comprises a sample acquisition module 1201 and a data processing module, wherein the sample acquisition module 1201 is used for acquiring a ground surface sample image and an initial ground surface coverage image corresponding to the ground surface sample image, and each pixel point in the initial ground surface coverage image corresponds to initial label information indicating the type of a ground surface coverage;
a label determining module 1202, configured to determine, based on the initial label information corresponding to each pixel point in the initial earth surface coverage image and the earth surface sample image, target label information corresponding to each pixel point in the earth surface sample image;
a probability prediction module 1203, configured to predict, based on the surface sample image and the first neural network, a first probability that each pixel point in the surface sample image belongs to each of multiple preset surface covering categories;
the parameter adjusting module 1204 is configured to adjust a network parameter value in the first neural network based on the first probability and the target label information corresponding to each pixel point, so as to obtain a target neural network.
In one possible implementation, the tag determining module 1202 when configured to determine the target tag information corresponding to each pixel point in the surface sample image based on the initial tag information corresponding to each pixel point in the initial surface coverage image and the surface sample image, includes:
predicting a second probability that each pixel point in the surface sample image belongs to each surface covering category based on the surface sample image and a second neural network to be trained;
adjusting network parameter values in the second neural network based on the second probability and the initial label information corresponding to each pixel point to obtain an initial second neural network;
and determining target label information corresponding to each pixel point in the surface sample image based on the surface sample image, the initial second neural network and the initial label information corresponding to each pixel point.
In one possible implementation, the tag determining module 1202 when configured to determine the target tag information corresponding to each pixel point in the surface sample image based on the surface sample image, the initial second neural network and the initial tag information corresponding to each pixel point, includes:
inputting the earth surface sample image into a current second neural network to obtain a second probability that each pixel point in the earth surface sample image predicted at the current time belongs to each earth surface covering object category; the current second neural network is an initial second neural network or a neural network obtained by adjusting the initial second neural network at least once;
determining current adjusted label information corresponding to each pixel point in the earth surface sample image based on a second probability that each pixel point in the current predicted earth surface sample image belongs to each earth surface covering category;
adjusting network parameter values in the current second neural network based on the second probability corresponding to each pixel point predicted at the current time, the initial label information corresponding to each pixel point and the label information corresponding to each pixel point after the last adjustment to obtain an adjusted current second neural network;
and under the condition that the preset cut-off condition is not met, returning to the step of inputting the earth surface sample image into the current second neural network until the preset cut-off condition is met, and taking the label information of each pixel point after the current adjustment as the target label information corresponding to the pixel point.
In one possible embodiment, the preset cutoff condition includes:
at least one of the current adjustment times reaches the set time threshold and the loss value corresponding to the current second neural network is smaller than the set threshold, and,
target pixel points with more than a set number exist in the earth surface sample image; and the corresponding label information after the current adjustment of the target pixel point is the same as the corresponding label information after the last adjustment.
In one possible implementation, the tag determining module 1202 when configured to determine the current adjusted tag information corresponding to each pixel point in the surface sample image based on the second probability that each pixel point predicted at the current time belongs to each surface covering category, includes:
determining the confidence coefficient that each pixel point belongs to the target earth surface covering category based on the second probability that the pixel point respectively belongs to different earth surface covering categories aiming at each pixel point predicted at the current time; the target earth surface covering category is the earth surface covering category with the maximum corresponding probability;
and determining the currently adjusted label information corresponding to each pixel point in the earth surface sample image based on the confidence coefficient that each pixel point predicted at the current time belongs to the target earth surface covering object category and the confidence coefficient threshold value corresponding to the earth surface sample image.
In a possible embodiment, the tag determining module 1202 when configured to determine, for each pixel point of the current prediction, a confidence that the pixel point belongs to the target earth surface covering category based on the second probabilities that the pixel point respectively belongs to different earth surface covering categories, includes:
acquiring the maximum probability and the second maximum probability in the second probability that the pixel point belongs to each earth surface covering category;
and taking the ratio of the maximum probability to the second maximum probability as the confidence coefficient that the pixel point belongs to the target earth surface covering object category.
In one possible implementation, the tag determination module 1202 is configured to determine the confidence threshold according to the following steps:
calculating a confidence coefficient mean value of each pixel point in the earth surface sample image belonging to the category of the target earth surface covering;
taking the confidence coefficient mean value as a confidence coefficient threshold value under the condition that the confidence coefficient mean value is less than or equal to the set threshold value;
and when the confidence coefficient average value is larger than the set threshold value, taking the set threshold value as the confidence coefficient threshold value.
In a possible implementation manner, the tag determining module 1202, when the module is configured to adjust the network parameter value in the current second neural network based on the second probability corresponding to each pixel point predicted at the current time, the initial tag information corresponding to each pixel point, and the last adjusted tag information corresponding to each pixel point, so as to obtain an adjusted current second neural network, includes:
determining a cross entropy loss value based on a second probability corresponding to each pixel point predicted at the current time and label information corresponding to each pixel point after last adjustment; determining a symmetric cross entropy loss value based on a second probability corresponding to each pixel point predicted at the current time and initial label information corresponding to each pixel point; determining a minimum entropy loss value based on a second probability corresponding to each pixel point of the current prediction;
determining a total loss value corresponding to the current second neural network based on the cross entropy loss value, the symmetric cross entropy loss value and the minimized entropy loss value;
and adjusting the network parameter value in the current second neural network based on the total loss value to obtain the adjusted current second neural network.
The description of the processing flow of each module in the device and the interaction flow between the modules may refer to the related description in the above method embodiments, and will not be described in detail here.
Corresponding to the method for identifying a surface covering in fig. 2, an embodiment of the present disclosure further provides an electronic device 1300, and as shown in fig. 13, a schematic structural diagram of the electronic device 1300 provided in the embodiment of the present disclosure includes:
a processor 101, a memory 102, and a bus 103; the storage 102 is used for storing execution instructions and includes a memory 1021 and an external storage 1022; the memory 1021 is also called an internal memory, and is used for temporarily storing the operation data in the processor 101 and the data exchanged with the external storage 1022 such as a hard disk, the processor 101 exchanges data with the external storage 1022 through the memory 1021, and when the electronic device 1300 is operated, the processor 101 communicates with the storage 102 through the bus 103, so that the processor 101 executes the following instructions: acquiring a ground surface image; classifying and identifying pixel point levels of the earth surface image by using a target neural network to obtain the category information of an earth surface covering to which the pixel points in the earth surface image belong; the target neural network is obtained by training through the earth surface sample image and target label information corresponding to sample pixel points in the earth surface sample image, the target label information is used for identifying the earth surface covering object type to which the sample pixel points belong, and the target label information is obtained by adjusting initial label information corresponding to the sample pixel points in the earth surface sample image.
Corresponding to the method for identifying a surface covering in fig. 4, an embodiment of the present disclosure further provides an electronic device 1400, as shown in fig. 14, a schematic structural diagram of the electronic device 1400 provided in the embodiment of the present disclosure includes:
a processor 111, a memory 112, and a bus 113; the storage 112 is used for storing execution instructions and includes a memory 1121 and an external storage 1122; the memory 1121 is also referred to as an internal memory, and is used for temporarily storing operation data in the processor 111 and data exchanged with the external memory 1122 such as a hard disk, the processor 111 exchanges data with the external memory 1122 through the memory 1121, and when the electronic device 1400 is operated, the processor 111 communicates with the memory 112 through the bus 113, so that the processor 111 executes the following instructions: acquiring a ground surface sample image and an initial ground surface covering image corresponding to the ground surface sample image, wherein each pixel point in the initial ground surface covering image corresponds to initial label information indicating the type of a ground surface covering; determining target label information corresponding to each pixel point in the earth surface sample image based on the initial label information corresponding to each pixel point in the initial earth surface covering image and the earth surface sample image; predicting a first probability that each pixel point in the earth surface sample image belongs to each earth surface covering category in multiple preset earth surface covering categories based on the earth surface sample image and the first neural network; and adjusting the network parameter value in the first neural network based on the first probability and the target label information corresponding to each pixel point to obtain the target neural network.
Embodiments of the present disclosure also provide a computer-readable storage medium having stored thereon a computer program, which, when being executed by a processor, performs the steps of the surface covering identification method described in the above method embodiments or performs the steps of the training method of the neural network. The storage medium may be a volatile or non-volatile computer-readable storage medium.
The computer program product of the surface covering identification method or the neural network training method provided in the embodiments of the present disclosure includes a computer-readable storage medium storing a program code, where instructions included in the program code may be used to execute steps of the surface covering identification method or the neural network training method described in the above method embodiments, which may be referred to in the above method embodiments specifically, and are not described herein again.
The embodiments of the present disclosure also provide a computer program, which when executed by a processor implements any one of the methods of the foregoing embodiments. The computer program product may be embodied in hardware, software or a combination thereof. In an alternative embodiment, the computer program product is embodied in a computer storage medium, and in another alternative embodiment, the computer program product is embodied in a Software product, such as a Software Development Kit (SDK), or the like.
It is clear to those skilled in the art that, for convenience and brevity of description, the specific working processes of the system and the apparatus described above may refer to the corresponding processes in the foregoing method embodiments, and are not described herein again. In the several embodiments provided in the present disclosure, it should be understood that the disclosed system, apparatus, and method may be implemented in other ways. The above-described embodiments of the apparatus are merely illustrative, and for example, the division of the units is only one logical division, and there may be other divisions when actually implemented, and for example, a plurality of units or components may be combined or integrated into another system, or some features may be omitted, or not executed. In addition, the shown or discussed mutual coupling or direct coupling or communication connection may be an indirect coupling or communication connection of devices or units through some communication interfaces, and may be in an electrical, mechanical or other form.
The units described as separate parts may or may not be physically separate, and parts displayed as units may or may not be physical units, may be located in one place, or may also be distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of the embodiment.
In addition, functional units in the embodiments of the present disclosure may be integrated into one processing unit, or each unit may exist alone physically, or two or more units are integrated into one unit.
The functions, if implemented in the form of software functional units and sold or used as a stand-alone product, may be stored in a non-volatile computer-readable storage medium executable by a processor. Based on such understanding, the technical solution of the present disclosure may be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device) to execute all or part of the steps of the method according to the embodiments of the present disclosure. And the aforementioned storage medium includes: various media capable of storing program codes, such as a usb disk, a removable hard disk, a Read-only Memory (ROM), a Random Access Memory (RAM), a magnetic disk, or an optical disk.
Finally, it should be noted that: the above-mentioned embodiments are merely specific embodiments of the present disclosure, which are used for illustrating the technical solutions of the present disclosure and not for limiting the same, and the scope of the present disclosure is not limited thereto, and although the present disclosure is described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: any person skilled in the art can modify or easily conceive of the technical solutions described in the foregoing embodiments or equivalent technical features thereof within the technical scope of the present disclosure; such modifications, changes or substitutions do not depart from the spirit and scope of the embodiments of the present disclosure, and should be construed as being included therein. Therefore, the protection scope of the present disclosure shall be subject to the protection scope of the claims.
Claims (19)
1. A method of identifying a surface covering, comprising:
acquiring a ground surface image;
carrying out classification and identification of pixel point levels on the earth surface image by using a target neural network to obtain the category information of an earth surface covering to which the pixel points in the earth surface image belong;
the target neural network is obtained by training through a surface sample image and target label information corresponding to sample pixel points in the surface sample image, the target label information is used for identifying the surface covering object type to which the sample pixel points belong, and the target label information is obtained by adjusting initial label information corresponding to the sample pixel points in the surface sample image.
2. The method of claim 1, wherein after obtaining the classification information of the surface covering to which the pixel points in the surface image belong, the method further comprises:
determining the display color of the pixel point in the earth surface covering image corresponding to the earth surface image based on the category information corresponding to each pixel point in the earth surface image;
and generating a ground surface covering image corresponding to the ground surface image based on the display color corresponding to each pixel point.
3. A method of training a neural network, comprising:
acquiring a ground surface sample image and an initial ground surface coverage image corresponding to the ground surface sample image, wherein each pixel point in the initial ground surface coverage image corresponds to initial label information indicating the ground surface coverage object type;
determining target label information corresponding to each pixel point in the earth surface sample image based on the initial label information corresponding to each pixel point in the initial earth surface covering image and the earth surface sample image;
predicting a first probability that each pixel point in the surface sample image belongs to each of a plurality of preset surface covering categories based on the surface sample image and a first neural network;
and adjusting the network parameter value in the first neural network based on the first probability corresponding to each pixel point and the target label information to obtain a target neural network.
4. The training method according to claim 3, wherein the determining target tag information corresponding to each pixel point in the surface sample image based on the initial tag information corresponding to each pixel point in the initial surface coverage image and the surface sample image comprises:
predicting a second probability that each pixel point in the surface sample image belongs to each surface covering category based on the surface sample image and a second neural network to be trained;
adjusting network parameter values in the second neural network based on the second probability corresponding to each pixel point and the initial label information to obtain an initial second neural network;
and determining target label information corresponding to each pixel point in the surface sample image based on the surface sample image, the initial second neural network and the initial label information corresponding to each pixel point.
5. The training method of claim 4, wherein the determining target label information corresponding to each pixel point in the surface sample image based on the surface sample image, the initial second neural network and the initial label information corresponding to each pixel point comprises:
inputting the earth surface sample image into a current second neural network to obtain a second probability that each pixel point in the earth surface sample image predicted at the current time belongs to each earth surface covering category; the current second neural network is the initial second neural network or a neural network obtained by adjusting the initial second neural network at least once;
determining current adjusted label information corresponding to each pixel point in the earth surface sample image based on a second probability that each pixel point in the current predicted earth surface sample image belongs to each earth surface covering category;
adjusting the network parameter values in the current second neural network based on the second probability corresponding to each pixel point predicted at the current time, the initial label information corresponding to each pixel point and the label information corresponding to each pixel point after the last adjustment to obtain an adjusted current second neural network;
and under the condition that a preset cut-off condition is not met, returning to the step of inputting the earth surface sample image into the current second neural network, and taking the label information of each pixel point after the current adjustment as the target label information corresponding to the pixel point after the preset cut-off condition is met.
6. Training method according to claim 5, characterized in that said preset cut-off condition comprises:
at least one of the current adjustment times reaches the set time threshold and the loss value corresponding to the current second neural network is smaller than the set threshold, and,
target pixel points with more than a set number exist in the earth surface sample image; and the corresponding label information after the current adjustment of the target pixel point is the same as the corresponding label information after the last adjustment.
7. The training method of claim 5, wherein determining the current adjusted label information corresponding to each pixel point in the surface sample image based on the second probability that each pixel point predicted at the current time belongs to each surface covering category comprises:
determining the confidence coefficient that each pixel point belongs to the target earth surface covering category based on the second probability that the pixel point respectively belongs to different earth surface covering categories aiming at each pixel point predicted at the current time; the target earth surface covering category is the earth surface covering category with the maximum corresponding probability;
and determining the current adjusted label information corresponding to each pixel point in the earth surface sample image based on the confidence coefficient of each pixel point of the current prediction belonging to the category of the target earth surface covering and the confidence coefficient threshold corresponding to the earth surface sample image.
8. The training method of claim 7, wherein the determining, for each pixel of the current prediction, the confidence that the pixel belongs to the target earth surface covering category based on the second probabilities that the pixel belongs to different earth surface covering categories respectively comprises:
acquiring the maximum probability and the second maximum probability in the second probability that the pixel point belongs to each earth surface covering category;
and taking the ratio of the maximum probability to the second maximum probability as the confidence coefficient that the pixel point belongs to the category of the target earth surface covering object.
9. Training method according to claim 7 or 8, wherein the confidence threshold is determined according to the following steps:
calculating the confidence coefficient mean value of each pixel point in the earth surface sample image belonging to the category of the target earth surface covering;
taking the confidence coefficient mean value as the confidence coefficient threshold value when the confidence coefficient mean value is less than or equal to a set threshold value;
and taking the set threshold as the confidence coefficient threshold when the confidence coefficient average value is larger than the set threshold.
10. The training method according to claim 5, wherein the adjusting the network parameter values in the current second neural network based on the second probability corresponding to each pixel point predicted at the current time, the initial label information corresponding to each pixel point, and the label information corresponding to each pixel point after the last adjustment to obtain the adjusted current second neural network comprises:
determining a cross entropy loss value based on the second probability corresponding to each pixel point predicted at the current time and the label information corresponding to each pixel point after the last adjustment; determining a symmetric cross entropy loss value based on the second probability corresponding to each pixel point predicted at the current time and the initial label information corresponding to each pixel point; and determining a minimum entropy loss value based on the second probability corresponding to each pixel point of the current prediction;
determining a total loss value corresponding to the current second neural network based on the cross entropy loss value, the symmetric cross entropy loss value and the minimized entropy loss value;
and adjusting the network parameter value in the current second neural network based on the total loss value to obtain the adjusted current second neural network.
11. A surface covering identification arrangement, comprising:
the image acquisition module is used for acquiring a ground surface image;
the category identification module is used for carrying out pixel point level classification identification on the earth surface image by using a target neural network to obtain category information of an earth surface covering object to which the pixel points in the earth surface image belong;
the target neural network is obtained by training through a surface sample image and target label information corresponding to sample pixel points in the surface sample image, the target label information is used for identifying the surface covering object type to which the sample pixel points belong, and the target label information is obtained by adjusting initial label information corresponding to the sample pixel points in the surface sample image.
12. The earth surface covering recognition device of claim 11, further comprising an image generation module, wherein after the category video module obtains the category information of the earth surface covering to which the pixel points in the earth surface image belong, the image generation module is configured to:
determining the display color of the pixel point in the earth surface covering image corresponding to the earth surface image based on the category information corresponding to each pixel point in the earth surface image;
and generating a ground surface covering image corresponding to the ground surface image based on the display color corresponding to each pixel point.
13. An apparatus for training a neural network, comprising:
the system comprises a sample acquisition module, a data acquisition module and a data processing module, wherein the sample acquisition module is used for acquiring a ground surface sample image and an initial ground surface coverage image corresponding to the ground surface sample image, and each pixel point in the initial ground surface coverage image corresponds to initial label information indicating the type of a ground surface coverage object;
the label determining module is used for determining target label information corresponding to each pixel point in the earth surface sample image based on the initial label information corresponding to each pixel point in the initial earth surface coverage image and the earth surface sample image;
the probability prediction module is used for predicting a first probability that each pixel point in the earth surface sample image belongs to each earth surface covering category in multiple preset earth surface covering categories based on the earth surface sample image and the first neural network;
and the parameter adjusting module is used for adjusting the network parameter value in the first neural network based on the first probability corresponding to each pixel point and the target label information to obtain a target neural network.
14. The training apparatus of claim 13, wherein the label determination module, when configured to determine the target label information corresponding to each pixel point in the surface sample image based on the initial label information corresponding to each pixel point in the initial surface coverage image and the surface sample image, comprises:
predicting a second probability that each pixel point in the surface sample image belongs to each surface covering category based on the surface sample image and a second neural network to be trained;
adjusting network parameter values in the second neural network based on the second probability corresponding to each pixel point and the initial label information to obtain an initial second neural network;
and determining target label information corresponding to each pixel point in the surface sample image based on the surface sample image, the initial second neural network and the initial label information corresponding to each pixel point.
15. The training apparatus of claim 14, wherein the label determination module, when configured to determine the target label information corresponding to each pixel point in the surface sample image based on the initial label information corresponding to the surface sample image, the initial second neural network and each pixel point, comprises:
inputting the earth surface sample image into a current second neural network to obtain a second probability that each pixel point in the earth surface sample image predicted at the current time belongs to each earth surface covering category; the current second neural network is the initial second neural network or a neural network obtained by adjusting the initial second neural network at least once;
determining current adjusted label information corresponding to each pixel point in the earth surface sample image based on a second probability that each pixel point in the current predicted earth surface sample image belongs to each earth surface covering category;
adjusting the network parameter values in the current second neural network based on the second probability corresponding to each pixel point predicted at the current time, the initial label information corresponding to each pixel point and the label information corresponding to each pixel point after the last adjustment to obtain an adjusted current second neural network;
and under the condition that a preset cut-off condition is not met, returning to the step of inputting the earth surface sample image into the current second neural network, and taking the label information of each pixel point after the current adjustment as the target label information corresponding to the pixel point after the preset cut-off condition is met.
16. Training apparatus according to claim 15, wherein the preset cut-off condition comprises:
at least one of the current adjustment times reaches the set time threshold and the loss value corresponding to the current second neural network is smaller than the set threshold, and,
target pixel points with more than a set number exist in the earth surface sample image; and the corresponding label information after the current adjustment of the target pixel point is the same as the corresponding label information after the last adjustment.
17. The training apparatus of claim 15, wherein the label determination module, when configured to determine the current adjusted label information corresponding to each pixel point in the surface sample image based on the second probability that each pixel point of the current prediction belongs to each surface covering category, comprises:
determining the confidence coefficient that each pixel point belongs to the target earth surface covering category based on the second probability that the pixel point respectively belongs to different earth surface covering categories aiming at each pixel point predicted at the current time; the target earth surface covering category is the earth surface covering category with the maximum corresponding probability;
and determining the current adjusted label information corresponding to each pixel point in the earth surface sample image based on the confidence coefficient of each pixel point of the current prediction belonging to the category of the target earth surface covering and the confidence coefficient threshold corresponding to the earth surface sample image.
18. An electronic device, comprising: a processor, a memory and a bus, the memory storing machine-readable instructions executable by the processor, the processor and the memory communicating over the bus when the electronic device is operating, the machine-readable instructions when executed by the processor performing the steps of the method of any of claims 1 to 10.
19. A computer-readable storage medium, having stored thereon a computer program which, when being executed by a processor, is adapted to carry out the steps of the method according to any one of claims 1 to 10.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010606948.7A CN111753773A (en) | 2020-06-29 | 2020-06-29 | Surface covering object recognition method, and neural network training method and device |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010606948.7A CN111753773A (en) | 2020-06-29 | 2020-06-29 | Surface covering object recognition method, and neural network training method and device |
Publications (1)
Publication Number | Publication Date |
---|---|
CN111753773A true CN111753773A (en) | 2020-10-09 |
Family
ID=72678121
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202010606948.7A Withdrawn CN111753773A (en) | 2020-06-29 | 2020-06-29 | Surface covering object recognition method, and neural network training method and device |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN111753773A (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112802108A (en) * | 2021-02-07 | 2021-05-14 | 上海商汤科技开发有限公司 | Target object positioning method and device, electronic equipment and readable storage medium |
-
2020
- 2020-06-29 CN CN202010606948.7A patent/CN111753773A/en not_active Withdrawn
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112802108A (en) * | 2021-02-07 | 2021-05-14 | 上海商汤科技开发有限公司 | Target object positioning method and device, electronic equipment and readable storage medium |
CN112802108B (en) * | 2021-02-07 | 2024-03-15 | 上海商汤科技开发有限公司 | Target object positioning method, target object positioning device, electronic equipment and readable storage medium |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107092870B (en) | A kind of high resolution image Semantic features extraction method | |
Awad et al. | Multicomponent image segmentation using a genetic algorithm and artificial neural network | |
Tolpekin et al. | Quantification of the effects of land-cover-class spectral separability on the accuracy of Markov-random-field-based superresolution mapping | |
Su et al. | Combining Hopfield neural network and contouring methods to enhance super-resolution mapping | |
Li et al. | A spatial–temporal Hopfield neural network approach for super-resolution land cover mapping with multi-temporal different resolution remotely sensed images | |
CN113449594A (en) | Multilayer network combined remote sensing image ground semantic segmentation and area calculation method | |
Li et al. | Using a sub-pixel mapping model to improve the accuracy of landscape pattern indices | |
Sales et al. | Land use and land cover area estimates from class membership probability of a random forest classification | |
CN109685806B (en) | Image significance detection method and device | |
CN107506792B (en) | Semi-supervised salient object detection method | |
Ling et al. | Learning-based superresolution land cover mapping | |
CN108428220A (en) | Satellite sequence remote sensing image sea island reef region automatic geometric correction method | |
CN111144234A (en) | Video SAR target detection method based on deep learning | |
CN112669298A (en) | Foundation cloud image cloud detection method based on model self-training | |
CN104077765B (en) | Image segmentation device, image partition method | |
Li et al. | Cropland data fusion and correction using spatial analysis techniques and the Google Earth Engine | |
Pan et al. | An adaptive multifeature method for semiautomatic road extraction from high-resolution stereo mapping satellite images | |
CN115239698A (en) | Change detection method and system based on multi-level feature fusion of subdivision grid images | |
Nowosad et al. | Extended SLIC superpixels algorithm for applications to non-imagery geospatial rasters | |
CN116343048A (en) | Accurate land block boundary extraction method and system for plain crop type complex region | |
CN112131969A (en) | Remote sensing image change detection method based on full convolution neural network | |
CN111753773A (en) | Surface covering object recognition method, and neural network training method and device | |
CN110084820B (en) | Purple soil image self-adaptive segmentation extraction method based on improved FCM algorithm | |
Zhang et al. | Spectral–spatial based sub-pixel mapping of remotely sensed imagery with multi-scale spatial dependence | |
CN112329818A (en) | Hyperspectral image unsupervised classification method based on graph convolution network embedded representation |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
WW01 | Invention patent application withdrawn after publication | ||
WW01 | Invention patent application withdrawn after publication |
Application publication date: 20201009 |