CN114078237A - Remote sensing image road change identification method and device - Google Patents
Remote sensing image road change identification method and device Download PDFInfo
- Publication number
- CN114078237A CN114078237A CN202111400764.6A CN202111400764A CN114078237A CN 114078237 A CN114078237 A CN 114078237A CN 202111400764 A CN202111400764 A CN 202111400764A CN 114078237 A CN114078237 A CN 114078237A
- Authority
- CN
- China
- Prior art keywords
- road
- remote sensing
- training
- model
- road change
- 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.)
- Pending
Links
Images
Classifications
-
- 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
- G06F18/2155—Generating training patterns; Bootstrap methods, e.g. bagging or boosting characterised by the incorporation of unlabelled data, e.g. multiple instance learning [MIL], semi-supervised techniques using expectation-maximisation [EM] or naïve labelling
-
- 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
-
- 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
- G06N3/08—Learning methods
- G06N3/088—Non-supervised learning, e.g. competitive learning
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Image Analysis (AREA)
- Traffic Control Systems (AREA)
Abstract
The invention discloses a method and a device for identifying road changes of remote sensing images. The method comprises the following steps: acquiring new and old remote sensing images of unmarked roads; cutting the remote sensing image; carrying out contrast processing on the cut road remote sensing image; inputting the road change remote sensing image after the comparison processing into a road change recognition model for self-supervision training to obtain a pre-training model; training a pre-training model by adopting a road remote sensing image with a label to obtain a classification recognition model; and inputting the road change remote sensing image to be identified into the classification identification model of the migration self-supervision model parameter to obtain a road change result. The method provided by the invention can intelligently identify the road change by using the advantages of no need of manual marking, high efficiency and less expenditure, realizes high-efficiency monitoring and dynamic supervision of the road change condition, and provides technical and data support for promoting the maintenance and optimization of traffic roads and implementing the detection and treatment of road hidden dangers.
Description
Technical Field
The invention relates to the technical field of geographic information remote sensing data analysis, in particular to a road change identification method and device.
Background
With the rapid development of industrialization and urbanization in China, the number of roads is rapidly increased. Particularly, in recent years, the urban scale of our country is rapidly expanded, and the phenomenon of disordered expansion occurs, so that many new roads and many old roads disappear, and the original roads are greatly changed in width bifurcation and the like. Marking road changes requires a large amount of labor, and hinders real-time monitoring of road conditions.
In recent years, with the help of the development of artificial intelligence, a phenomenon that computers replace manpower appears in many fields, a satisfactory result is obtained in the aspect of remote sensing image classification, and the method is widely applied to the fields of urban construction planning identification and the like. However, the existing artificial intelligence recognition technology has great demand for highly accurate labeling, and the labeling is changed along with the change of the road along with the change of the time, so that the long-term and efficient recognition of the road change cannot be met.
For the problem of insufficient labeled samples, the common method is to adopt various data enhancement technologies to generate more samples, which indeed improves the effect of the model and is widely used, but the improvement degree is not enough to be practically applied; in addition, the technical development of the transfer learning slightly alleviates the problems that the model is pre-trained on a larger training set related to the current task, and the parameters of the pre-trained model are initialized to the parameters of the existing task model, but no large-scale remote sensing image training set exists at present, which brings difficulty to the application of the transfer learning.
Considering that the number of labeled remote sensing images is small, but the data of unlabeled remote sensing images is very rich in the world, and how to utilize the unprocessed massive data is a key thing, which is also a thing that can be done by the self-supervision learning. The self-supervision learning does not depend on any marked image data, the image is directly adopted to design an auxiliary task, and then higher-order and more complex feature knowledge of the image is learned, and the image is applied to a downstream task, so that the task effect is improved. At present, there are also many researches on processing application of label-free remote sensing image data.
The patent of publication CN112766089A discloses a cross-domain road extraction method based on a global-local countermeasure learning framework, which includes the following steps: step 1, acquiring source domain data with labels and target domain images without labels, forming a training set together for network training, and carrying out normalization and data enhancement on the training set; step 2, constructing a global-local countermeasure learning framework; step 3, training on a training set based on the constructed global-local countermeasure learning framework, and optimizing model parameters until convergence; and 4, predicting the test set from the target domain based on the converged global-local countermeasure learning framework, and obtaining a road extraction result by using the output road segmentation probability map. The implementation manner of the step 3 is as follows, step 3.1, initializing network parameters of a global-local antagonistic learning frame, fixing the cutting size of a training set, the input number of each batch, and the initial segmentation learning rate and the initial antagonistic learning rate; and 3.2, training by adopting an SGD (generalized decision device) optimizer, training by adopting an Adam optimizer, alternately performing segmentation training and countermeasure training step by step, and continuously updating network parameters of a global-local countermeasure learning framework through forward propagation and backward feedback until convergence.
The patent discloses that although a labeled and unlabeled remote sensing image is adopted to construct a road recognition model, a framework which adopts counterstudy only needs a large number of positive and negative samples to perform counterstudy, a large amount of manual work is consumed to go to a mobile phone to train labeled data, and only the framework of a road can be extracted and processed. The invention researches the self-supervision learning method, trains a large amount of label-free numbers to construct the recognition model by adopting the self-supervision comparison learning method without any manpower, and updates parameters by adopting a momentum method, so that the training speed is higher and the efficiency is higher.
The patent with the publication number of CN113486827A discloses a multi-source remote sensing image transfer learning method based on domain confrontation and self-supervision, a deep learning model is trained in a transfer learning frame based on domain confrontation training and self-supervision training after being pre-trained in a source domain, remote sensing image data of a target domain is input into the trained deep learning model, and a prediction result of the target domain is output; the transfer learning framework comprises a domain confrontation training module and a self-supervision training module; the domain confrontation training module comprises 1 generator and more than 1 discriminator; the self-supervision training module comprises an information entropy calculation module and a pseudo label selection module, the information entropy calculation module is used for calculating the information entropy of each sample prediction result, and the pseudo label selection module is used for selecting the pseudo labels according to the information entropy of each sample prediction result in an ordering mode.
Although the patent discloses that the remote sensing image transfer learning model is optimized by adopting the self-supervision training method, the self-supervision training method focuses more on the fine-tuning operation of the transfer learning, and neglects the strong pre-training capability of the self-supervision learning. According to the invention, through research on the self-supervision learning, the MoCo self-supervision comparison learning method is adopted to carry out comparison learning training on the unlabelled data, so that a pre-training model suitable for downstream tasks in various fields is obtained, the transfer capability is strong in the field of remote sensing images, and the pre-training model has practical significance when being applied to a specific task of remote sensing image road change identification.
Disclosure of Invention
The invention provides a method and a device for identifying road changes of remote sensing images by using an artificial intelligence technology of self-supervision contrast learning in order to solve the defects of the prior art; the method is realized by the following technical scheme:
the invention aims to provide a remote sensing image road change identification method, which comprises the following steps:
acquiring a large number of new remote sensing images and old remote sensing images of unmarked roads;
cutting the new and old remote sensing images according to the plot vector range corresponding to the road to be identified to obtain new and old cut remote sensing images containing the plot vector range;
comparing the remote sensing images of the new and old cut roads to obtain the remote sensing images of the new and old roads;
carrying out self-supervision training on the remote sensing images of the new road and the old road to obtain a pre-training model;
training a pre-training model by adopting a small number of road change remote sensing image samples with labels to obtain a classification recognition model of migration self-supervision model parameters;
and inputting the road change remote sensing image to be identified into the classification identification model of the migration self-supervision model parameter, and acquiring the road change result corresponding to the road to be identified, which is output by the classification identification model.
Preferably, in the process of processing data, the comparison processing is to perform special subtraction on the new road remote sensing image and the old road remote sensing image, wherein the new image has a road area but the old image has no 1, the new image has no road area but the old image has no 2, and the rest conditions are all 0.
Preferably, the self-supervised training: the deep learning model is one or more of VGG16, ResNet18, ResNet50 or DenseNet 121; training by adopting a comparison-based self-supervision learning method, namely training an encoder in the training process, and ensuring that an output vector of the encoder is similar to a corresponding key and is not similar to other keys in a large dictionary; the loss function adopts InfonCE, and the effect is that when the query vector q is similar to a positive key value k and is not similar to a negative key value k, the corresponding loss is reduced, and a specific formula is as follows:
where K is the number of negative key values and τ is a hyperparameter.
Preferably, in the comparison-based self-supervision learning method, the dictionary uses a momentum comparison (MoCo) method, namely, a queue is used for storing the dictionary; after each new batch is coded, entering a queue, and then discharging the oldest code from the queue;
preferably, the remote sensing images of the new road and the old road are subjected to self-supervision training, and the method comprises the following specific steps:
carrying out data enhancement processing on each remote sensing image to obtain a query graph and a template graph;
inputting the data into a deep learning model to respectively extract query features and template features;
setting the template features without using gradient update parameters, assuming the template features are stable;
calculating the matching degree of the features by using matrix multiplication;
setting an NCE loss function to ensure that the matching degree of the NCE and the pictures derived by the NCE is maximum;
updating parameters of the query network by using a gradient descent method according to the loss, and updating parameters of the template network by using a momentum method;
and repeating the steps to finish the training of the self-supervision contrast learning, and obtaining a pre-training model.
Preferably, the training method of the classification recognition model of the migration self-supervision model parameters specifically includes:
obtaining a road change remote sensing image sample with a label and a road change type contained in the sample;
reserving parameters of a pre-training model, accessing a linear classification network layer behind an encoder, and identifying the road change condition;
training by adopting a road change remote sensing image sample with a label, and finely adjusting parameters of the model to obtain a training model;
inputting the road change remote sensing image sample with the label into a training model, and obtaining a predicted road change type output by the training model;
calculating to obtain a training loss value according to the road change type and the predicted road change type, wherein a loss function is as follows:
where y is the desired output and a is the output of the neuron.
And under the condition that the loss value is less than 0.1, taking the training model as a final road change recognition model, namely a classification recognition model for migrating the self-supervision model parameters.
The invention also aims to provide a remote sensing image road change recognition device which comprises a road change acquisition module to be recognized, a cutting remote sensing image acquisition module, a self-supervision comparison learning characteristic processing module and a target recognition result acquisition module;
the road change acquisition module is used for acquiring new and old remote sensing images of the road to be identified;
the cutting remote sensing image acquisition module is used for cutting the new and old remote sensing images according to the plot vector range corresponding to the road to be identified to obtain the new and old cut remote sensing images only containing the road vector range;
the self-supervision contrast learning feature processing module is used for self-supervision training of the non-label cut remote sensing image to obtain a feature extractor containing specific parameters;
and the target identification result acquisition module is used for inputting the road change remote sensing image to be identified into the road change identification model and acquiring the road change type corresponding to the road change to be identified, which is output by the road change identification model.
Preferably, the remote sensing image road change recognition device further comprises a model training sample acquisition module, a plot remote sensing image generation module, a road change remote sensing image generation module, a self-supervision learning model parameter generation module, a predicted road change type acquisition module, a loss value calculation module and a road change type recognition module acquisition module;
the model training sample acquisition module is used for acquiring a model training sample; the model training sample comprises new and old road remote sensing images and road change types contained in the remote sensing images;
the land parcel remote sensing image generating module is used for cutting the new and old remote sensing images according to the road vector ranges of the new and old land parcels to generate new and old road cut remote sensing images corresponding to the land parcels;
the road change remote sensing image generation module is used for carrying out contrast processing on the cut remote sensing images of the new road and the old road to generate new road and old road change images corresponding to the plurality of plots;
the self-supervision learning model parameter generation module is used for carrying out self-supervision training on the new and old road change images to update model parameters, generating parameters for feature extraction of the previous layers of networks and obtaining a training model;
the predicted road change type obtaining module is used for inputting the road change remote sensing image to be recognized into a training model and obtaining the predicted road change types of the plurality of plots output by the training model;
the loss value calculation module is used for calculating the loss value of the training model according to the road change type and the predicted road change type;
and the road change type identification module acquisition module is used for taking the training model as a final road change identification model under the condition that the loss value is within a preset range.
Compared with the prior art, the invention has the advantages that:
the invention adopts the marker post models VGG16, ResNet18, ResNet50 or DenseNet121 in a plurality of image classification fields as the framework network for self-supervised learning to train, and the effect on image classification can reach the top level in the industry.
The method adopts an image contrast processing method to generate a contrast image of the road change, an image processing library of matlab is used for processing noise generated in the image subtraction process, holes appearing in the road change image are filled, finally, thresholds are set to assign 0, 1 and 2 to different blocks, and the coincidence rate of the final result and the manual mark is up to 99%.
The method adopts a MoCo self-supervision contrast learning method, can fully train information contained in the images, enables the images to mutually resist learning, and excavates high-order information in the images, so that an image classification model with a better effect is generated, the accuracy rate is improved by 20% in the task of road change recognition compared with that of the traditional technology, and the accuracy rate reaches 96%; and most data in the training process are initially label-free, no manual work is needed, the training speed is higher, and the efficiency is higher.
According to the method, the road change data are collected and sorted in time, the satellite remote sensing images are updated regularly, the self-supervision contrast learning, the transfer learning, the remote sensing image visual analysis and other technologies are utilized for analysis and processing, the road change current situation is rapidly and accurately acquired, the automation and the intelligence of the model are improved, manual marking of data is not needed, and rapid and accurate automatic identification of the road change types is achieved. And moreover, the change conditions of roads in each city are updated in time, the dynamic management level of the traffic planning city construction is comprehensively improved, and powerful data support and automation technology is provided for effective management and scientific decision of the city planning.
Drawings
FIG. 1 is a flow chart of a method for identifying road changes in remote sensing images according to the present invention;
FIG. 2 is a schematic structural diagram of a road change recognition device according to the present invention;
fig. 3 is a schematic diagram of an application system of the road change identification method according to embodiment 3 of the present invention.
Detailed Description
The technical solution of the present invention is further defined below with reference to the specific embodiments, but the scope of the claims is not limited to the description.
Example one
A remote sensing image road change identification method is shown in a flow chart of fig. 1 and comprises the following steps:
step 1: acquiring new and old remote sensing images of a large number of unmarked roads;
the remote sensing image of the road is a remote sensing image of a land parcel containing the road obtained by satellite shooting, and as the remote sensing image is updated along with time, the latest remote sensing image and the remote sensing image before one year are used for new-old comparison.
Step 2: cutting the new and old remote sensing images according to the plot vector range corresponding to the road to be identified to obtain new and old cut remote sensing images containing the plot vector range;
the block vector range refers to the area range of the block where the road is located.
And step 3: comparing the remote sensing images of the new and old cut roads to obtain the remote sensing images of the new and old roads;
the comparison treatment specifically comprises the following steps: in the process of processing data, a new road remote sensing image and an old road remote sensing image are subjected to special subtraction, the new image has a road area but the old image has no 1, the new image has no road area but the old image has no 2, and the rest conditions are all 0.
And 4, step 4: inputting the remote sensing images of the new road and the old road to a deep learning model for self-supervision training to obtain a pre-training model, and the method specifically comprises the following steps:
step (4.1): and performing data enhancement processing on each remote sensing image to obtain a query graph and a template graph.
In the embodiment, a road change database can be constructed in advance in consideration of more used remote sensing image data; the road change database contains information such as road traffic conditions, road change time, new and old remote sensing images of roads and the like.
Step (4.2): inputting the query features and the template features into a deep continuous learning model to respectively extract the query features and the template features;
in the model construction process, two networks of f _ k and f _ q are used for mapping input information to a feature space, wherein the feature space has a vector representation with the length of c, the f _ k network extracts the features of the template graph, and the f _ q network extracts the features of the query graph. The structures of the two networks are ResNet50, and can be changed into one or more of VGG16, ResNet18 and DenseNet121 according to requirements;
step (4.3): setting the template features without using gradient update parameters, assuming the template features are stable;
although f _ k and f _ q adopt the same network structure, the same network structure needs to be distinguished in the aspect of parameter updating, similarity matching is carried out on the query characteristic and the template characteristic in the training process, so that the template characteristic is assumed to be stable, and the parameters do not need to be quickly updated by a gradient algorithm;
step (4.4): calculating the matching degree of the features by using matrix multiplication;
step (4.5): setting an NCE loss function to ensure that the matching degree of the NCE and the pictures derived by the NCE is maximum;
step (4.6): updating parameters of the query network by using gradient interplanetary according to the loss, and updating parameters of the template network by using a momentum method;
and (4.1) repeating the steps (4.1) to (4.6) until the loss value of the InfonCE reaches the requirement or the number of training iterations reaches the requirement, and obtaining a pre-training model.
And 5: training a pre-training model by adopting a small number of labeled road change remote sensing image samples to obtain a classification recognition model of migration self-supervision model parameters, wherein the specific training method comprises the following steps:
step (5.1): obtaining a road change remote sensing image sample with a label and a road change type contained in the sample;
step (5.2): reserving parameters of a pre-trained deep learning model, accessing a linear classification network layer behind an encoder, and identifying the road change condition;
step (5.3): training by adopting a road change remote sensing image sample, and finely adjusting parameters of the model to obtain a training model;
step (5.4): inputting the road change remote sensing image sample with the label into a training model, and acquiring the predicted road change types of the plurality of plots output by the training model;
step (5.5): calculating a loss value according to the road change type and the predicted road change type, wherein a loss function is as follows:
where y is the desired output and a is the output of the neuron.
Step (5.6): and under the condition that the loss value is less than 0.1, taking the trained road change recognition model as a final road change recognition model, namely a classification recognition model of the migration self-supervision model parameters.
Step 6: and inputting the road change remote sensing image to be identified into the classification identification model of the migration self-supervision model parameter, and acquiring the road change result corresponding to the road to be identified, which is output by the classification identification model.
Example 2
A remote sensing image road change recognition device is shown in figure 2 and comprises the following modules:
the road to be identified change acquiring module 410 is used for acquiring new and old remote sensing images of the road to be identified;
the cutting remote sensing image obtaining module 420 is configured to cut the remote sensing image according to the parcel vector range corresponding to the road to be identified, so as to obtain a cut remote sensing image only including the road vector range;
the self-supervision comparison learning characteristic processing module 430 is used for self-supervision training of the non-label road change remote sensing image to obtain a characteristic extractor containing specific parameters;
and the target recognition result obtaining module 440 is configured to input the remote sensing image of the road change to be recognized into a road change recognition model trained in a fine tuning manner in advance, and obtain a road change type corresponding to the road change to be recognized, which is output by the road change recognition model.
Further, the remote sensing image road change recognition device further comprises:
the model training sample acquisition module is used for acquiring a model training sample; the sample comprises new and old road remote sensing images and road change types contained in the remote sensing images;
the land parcel remote sensing image generating module is used for cutting the new and old remote sensing images according to the road vector ranges of the new and old land parcels to generate new and old road cut remote sensing images corresponding to the land parcels;
the road change remote sensing image generation module is used for performing contrast processing on the remote sensing images cut according to the new road and the old road to generate road change images corresponding to the plurality of plots;
the self-supervision learning model parameter generation module is used for carrying out self-supervision training according to the road change image to update model parameters, generating parameters for feature extraction of the previous layers of networks and obtaining a training model;
the predicted road change type obtaining module is used for inputting the road change images into a training model and obtaining predicted road change types of the plurality of plots output by training;
the loss value calculation module is used for calculating the loss value of the training model according to the road change type and the predicted road change type;
and the road change type identification module acquisition module is used for taking the training model as a final road change identification model under the condition that the loss value is within a preset range.
Example 3
In this embodiment, the identification method in embodiment 1 and the identification device in embodiment 2 are applied to an actual road identification process to form a road identification change identification system, as shown in fig. 3, including a road change database, a sample training management module, a data preprocessing module, an intelligent identification module, and a human-computer interaction module.
Road change database
The road change database contains information such as road traffic conditions, road change time, new and old remote sensing images of roads and the like.
Second, data preprocessing module
(1) Markless remote sensing image data acquisition module
According to the existing large amount of unmarked remote sensing image data and the updating time, the data are imported into a road change database in batch without manual processing and marking, and become a data set of self-supervision training.
(2) Data extraction submodule
And the data extraction submodule stores the remote sensing images of the plots with the roads in the road change database in a temporary storage directory according to the plot names and then takes down the next plot image until all the plot images are taken.
(3) Image cutting submodule
And the data cutting submodule takes out the remote sensing image and the corresponding road vector data from the temporary storage catalogue for cutting, and the cut image is stored in the cut image catalogue.
(4) Road change image generation submodule
And the road change image generation submodule extracts and processes data from the cut image directory, and in the process of processing the data, a new road remote sensing image and an old road remote sensing image are used for carrying out special subtraction, wherein the new image has a road area but the old image has no 1, the new image has no road area but the old image has no 2, and the rest conditions are 0. And storing the processed image in an image directory to be identified.
Third, sample training management module
(1) Sample management submodule
The sample management submodule mainly realizes operations such as browsing and deleting of various samples and selects various samples participating in training.
(2) Sample editing tool
The sample editing tool mainly processes the marked images to be identified, and the number of the images is small, so that the processing is convenient. The system provides a plot sample editing tool, the main functions comprise functions of picture selection, sample modification, storage and uploading and the like, and the functions are described as follows:
selecting pictures: the pictures can be selected or uploaded from the database, and the operation can be carried out independently or in batches by using codes.
Sample modification: the selected samples may be manually modified.
Storing and uploading: and after the manual determination of the sample area is completed, the sample area needs to be saved, uploaded and updated to the server.
(3) Training task management submodule
After the samples are preprocessed, a self-supervision contrast learning component is required to be used for training to obtain a pre-training model, then a transfer learning technology and a remote sensing image visual analysis component are used for identifying and training sample characteristics, and the sample characteristics are extracted to form a sample library.
The training task management mainly realizes the functions of training sample selection, task starting and stopping, task monitoring, recognition rate checking and the like, and the functions are described as follows:
training sample selection: for selecting the samples to be involved in the training from all the samples of the pre-processing.
And (3) starting and stopping the task: starting self-supervision training after a training sample is selected; selecting a sample with a label, and then selecting a corresponding road change category to start deep learning training; the training may be terminated if it is necessary to stop the training for some reason during the training process.
And (3) task monitoring: the method is used for monitoring the training process, and information such as training progress, started time, predicted ending time and the like is displayed in the form of a progress bar.
Checking the recognition rate: after fine adjustment on the pre-training model is completed, one or more pictures can be manually selected for recognition, the manual inspection effect is quite in line with expectations, and if the effect is not in line with requirements, the sample can be added or replaced for re-training.
Fourth, intelligent recognition module
(1) Self-supervision contrast learning component
The method mainly comprises a basic framework of machine learning, a data batch acquisition and storage component, a GPU and CPU scheduling component, an automatic supervision structure component and the like.
(2) Transfer learning component
The method mainly comprises a basic framework of machine learning, an empirical value and fine tuning value reference component of a hyper-parameter, a real-time storage and deletion component of a training model and the like.
(3) Remote sensing image visual analysis component
And a neural network vision analysis algorithm model library aiming at remote sensing big data.
Five, man-machine interaction module
(1) Recognition result display submodule
The identification result display mainly comprises identification result display and query, and the identification result information comprises the change condition and the treatment state of the road and the like. The map positioning is supported, and the map information of the corresponding position can be directly positioned by clicking the picture.
(2) Task control submodule
And (3) each item and parcel service personnel in the image directory to be identified can stop, start, delete, set a monitoring interval and the like, if the system cannot find the required image at the monitoring time point, no image is prompted, and the service personnel is prompted to carry out on-site inspection.
(3) Manual determination submodule
The service personnel can identify the road change types automatically identified by the system, some roads can be reported according to the supervision requirements after the change types are confirmed, the system can verify the road change types which are difficult to accurately identify through remote sensing images, the service personnel can verify the road change types through on-site confirmation, and pictures checked on the site can be uploaded to the system to serve as reporting auxiliary materials.
The detailed description set forth herein may provide those skilled in the art with a more complete understanding of the present application, and is not intended to limit the present application in any way. Thus, it will be appreciated by those skilled in the art that modifications or equivalents may still be made to the present application; all technical solutions and modifications thereof which do not depart from the spirit and technical essence of the present application should be covered by the scope of protection of the present patent application.
Claims (8)
1. A method for identifying road changes of remote sensing images is characterized by comprising the following steps:
acquiring a large number of new remote sensing images and old remote sensing images of unmarked roads;
cutting the new and old remote sensing images according to the plot vector range corresponding to the road to be identified to obtain new and old cut remote sensing images containing the plot vector range;
comparing the remote sensing images of the new and old cut roads to obtain the remote sensing images of the new and old roads;
carrying out self-supervision training on the remote sensing images of the new road and the old road to obtain a pre-training model;
training a pre-training model by using a small amount of labeled road change remote sensing images to obtain a classification recognition model of migration self-supervision model parameters;
and inputting the road change remote sensing image to be identified into the classification identification model of the migration self-supervision model parameter, and acquiring the road change result corresponding to the road to be identified, which is output by the classification identification model.
2. The method for identifying road changes in remote-sensing images according to claim 1, wherein the comparison process is a special subtraction process using the remote-sensing images of the new road and the remote-sensing images of the old road during data processing, i.e. the new image has a road area but the old image has no 1, the new image has no road area but the old image has 2, and the rest is 0.
3. The method for identifying road changes in remote-sensing images as claimed in claim 1, wherein the self-supervision training comprises:
the deep learning model adopted by the self-supervision training is one or more of VGG16, ResNet18, ResNet50 or DenseNet 121;
training by adopting a comparison-based self-supervision learning method, namely training an encoder in the training process, and ensuring that an output vector of the encoder is similar to a corresponding key and is not similar to other keys in a large dictionary;
the loss function adopts InfonCE, and the effect is that when the query vector q is similar to a positive key value k and is not similar to a negative key value k, the corresponding loss is reduced, and a specific formula is as follows:
where K is the number of negative key values and τ is a hyperparameter.
4. The method for identifying road changes in remote sensing images as claimed in claim 3, wherein the comparison-based self-supervised learning method is a method in which a dictionary uses momentum comparison, i.e. a queue is used to store the dictionary; i.e. each new batch is encoded and then queued, and then the oldest is dequeued.
5. The method for identifying road changes in remote-sensing images as claimed in claim 1, wherein the method for performing self-supervision training on the remote-sensing images of new and old road changes comprises the following specific steps:
carrying out data enhancement processing on each remote sensing image to obtain a query graph and a template graph;
inputting the data into a deep learning model to respectively extract query features and template features;
setting the template features without using gradient update parameters, assuming the template features are stable;
calculating the matching degree of the features by using matrix multiplication;
setting an NCE loss function to ensure that the matching degree of the NCE and the pictures derived by the NCE is maximum;
updating parameters of the query network by using a gradient descent method according to the loss, and updating parameters of the template network by using a momentum method;
and repeating the steps to finish the training of the self-supervision contrast learning, and obtaining a pre-training model.
6. The method for identifying road changes in remote-sensing images as claimed in claim 1, wherein the training method of the classification identification model of the migration self-supervision model parameters specifically comprises:
obtaining a road change remote sensing image sample with a label and a road change type contained in the road change remote sensing image sample;
reserving parameters of a pre-training model, accessing a linear classification network layer behind an encoder, and identifying the road change condition;
training by adopting a road change remote sensing image sample with a label, and finely adjusting parameters of the model to obtain a training model;
inputting the road change remote sensing image sample with the label into a training model, and obtaining a predicted road change type output by the training model;
calculating to obtain a training loss value according to the road change type and the predicted road change type;
and under the condition that the loss value is within a preset range, taking the training model as a final road change recognition model, namely a classification recognition model for migrating the self-supervision model parameters.
7. The device for identifying the road change of the remote sensing image according to any one of claims 1 to 6, which is characterized by comprising a road change to be identified acquisition module, a cutting remote sensing image acquisition module, a self-supervision comparison learning characteristic processing module and a target identification result acquisition module;
the road change acquisition module is used for acquiring new and old remote sensing images of the road to be identified;
the cutting remote sensing image acquisition module is used for cutting the new and old remote sensing images according to the plot vector range corresponding to the road to be identified to obtain the new and old cut remote sensing images only containing the road vector range;
the self-supervision contrast learning feature processing module is used for self-supervision training of the non-label road change remote sensing image to obtain a feature extractor containing specific parameters;
and the target identification result acquisition module is used for inputting the road change remote sensing image to be identified into the road change identification model and acquiring the road change type corresponding to the road change to be identified, which is output by the road change identification model.
8. The image-sensing road change recognition device of claim 7, further comprising a model training sample acquisition module, a plot remote sensing image generation module, a road change remote sensing image generation module, a self-supervision learning model parameter generation module, a predicted road change type acquisition module, a loss value calculation module, and a road change type recognition module acquisition module;
the model training sample acquisition module is used for acquiring a model training sample; the model training sample comprises new and old road remote sensing images and road change types contained in the remote sensing images;
the land parcel remote sensing image generating module is used for cutting the new and old remote sensing images according to the road vector ranges of the new and old land parcels to generate new and old road cut remote sensing images corresponding to the land parcels;
the road change remote sensing image generation module is used for carrying out contrast processing on the cut remote sensing images of the new road and the old road to generate new road and old road change images corresponding to the plurality of plots;
the self-supervision learning model parameter generation module is used for carrying out self-supervision training on the changed images of the new road and the old road after comparison processing to update model parameters, generating parameters for feature extraction of the previous layers of networks and obtaining a training model;
the predicted road change type obtaining module is used for inputting the road change remote sensing image to be recognized into a training model and obtaining the predicted road change types of the plurality of plots output by the training model;
the loss value calculation module is used for calculating the loss value of the training model according to the road change type and the predicted road change type;
and the road change type identification module acquisition module is used for taking the training model as a final road change identification model under the condition that the loss value is within a preset range.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111400764.6A CN114078237A (en) | 2021-11-24 | 2021-11-24 | Remote sensing image road change identification method and device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111400764.6A CN114078237A (en) | 2021-11-24 | 2021-11-24 | Remote sensing image road change identification method and device |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114078237A true CN114078237A (en) | 2022-02-22 |
Family
ID=80284318
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111400764.6A Pending CN114078237A (en) | 2021-11-24 | 2021-11-24 | Remote sensing image road change identification method and device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114078237A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114863200A (en) * | 2022-03-21 | 2022-08-05 | 北京航空航天大学 | Gaze estimation method, device, and storage medium |
| CN114861865A (en) * | 2022-03-10 | 2022-08-05 | 长江三峡技术经济发展有限公司 | Self-supervision learning method, system, medium and electronic device of hyperspectral image classification model |
| WO2023207104A1 (en) * | 2022-04-26 | 2023-11-02 | 云南航天工程物探检测股份有限公司 | Ground penetrating radar tunnel lining quality inspection method based on self-supervised learning |
-
2021
- 2021-11-24 CN CN202111400764.6A patent/CN114078237A/en active Pending
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114861865A (en) * | 2022-03-10 | 2022-08-05 | 长江三峡技术经济发展有限公司 | Self-supervision learning method, system, medium and electronic device of hyperspectral image classification model |
| CN114863200A (en) * | 2022-03-21 | 2022-08-05 | 北京航空航天大学 | Gaze estimation method, device, and storage medium |
| WO2023207104A1 (en) * | 2022-04-26 | 2023-11-02 | 云南航天工程物探检测股份有限公司 | Ground penetrating radar tunnel lining quality inspection method based on self-supervised learning |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN114078237A (en) | Remote sensing image road change identification method and device | |
| CN114092832B (en) | High-resolution remote sensing image classification method based on parallel hybrid convolutional network | |
| CN110570651A (en) | Road network traffic situation prediction method and system based on deep learning | |
| CN106228808B (en) | City expressway travel time prediction method based on Floating Car space-time grid data | |
| CN110517482B (en) | Short-term traffic flow prediction method based on 3D convolutional neural network | |
| CN113222328B (en) | Air quality monitoring equipment point arrangement and site selection method based on road section pollution similarity | |
| CN109241227B (en) | Spatiotemporal data prediction modeling method based on stacking integrated learning algorithm | |
| CN110490249A (en) | Structural Damage Identification based on attribute class relations and few sample meta learning | |
| CN111127246A (en) | Intelligent prediction method for transmission line engineering cost | |
| CN111178628B (en) | Luggage arrival time prediction method and device | |
| CN112990558B (en) | Meteorological temperature and illumination prediction method based on deep migration learning | |
| CN106844636A (en) | A kind of unstructured data processing method based on deep learning | |
| CN114428828A (en) | Method and device for digging new road based on driving track and electronic equipment | |
| CN113158891B (en) | Cross-camera pedestrian re-identification method based on global feature matching | |
| CN111985325A (en) | Aerial small target rapid identification method in extra-high voltage environment evaluation | |
| CN116975256B (en) | Method and system for processing multisource information in construction process of underground factory building of pumped storage power station | |
| CN110032654A (en) | A kind of supermarket's commodity input method and system based on artificial intelligence | |
| CN114492978A (en) | Time-space sequence prediction method and device based on multi-layer attention mechanism | |
| CN117077005B (en) | Optimization method and system for urban micro-update potential | |
| CN114022688A (en) | Real-time livestock identity recognition method | |
| CN116244484B (en) | Federal cross-modal retrieval method and system for unbalanced data | |
| CN117078925A (en) | Building rubbish annual output accurate calculation method based on RDSA-deep LabV3+ network | |
| CN113487638A (en) | Ground feature edge detection method of high-precision semantic segmentation algorithm U2-net | |
| CN116108217B (en) | Fee evasion vehicle similar picture retrieval method based on depth hash coding and multitask prediction | |
| CN114822684B (en) | Method and system for predicting binding relationship of compound-target protein |
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 |