CN110008854A - Unmanned plane image Highway Geological Disaster recognition methods based on pre-training DCNN - Google Patents
Unmanned plane image Highway Geological Disaster recognition methods based on pre-training DCNN Download PDFInfo
- Publication number
- CN110008854A CN110008854A CN201910204723.6A CN201910204723A CN110008854A CN 110008854 A CN110008854 A CN 110008854A CN 201910204723 A CN201910204723 A CN 201910204723A CN 110008854 A CN110008854 A CN 110008854A
- Authority
- CN
- China
- Prior art keywords
- geological disaster
- image
- highway
- data
- training
- 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.)
- Granted
Links
Classifications
-
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/26—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion
- G06V10/267—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion by performing operations on regions, e.g. growing, shrinking or watersheds
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
- G06V20/182—Network patterns, e.g. roads or rivers
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- General Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The present invention relates to Highway Geological Disaster identification technology fields, disclose a kind of unmanned plane image Highway Geological Disaster recognition methods based on pre-training DCNN, the following steps are included: step S1, obtain the unmanned aerial vehicle remote sensing images within the scope of road extension, the orthography after absolute orientation is obtained after pretreatment；Step S2 divides pretreated unmanned aerial vehicle remote sensing images using the mean shift algorithm for considering image texture feature；Step S3 is applied in trained Highway Geological Disaster identification model using the unmanned aerial vehicle remote sensing images data of segmentation as input data, obtains Highway Geological Disaster recognition result.The present invention uses unmanned plane high resolution image, divide image based on the mean shift algorithm for considering textural characteristics, using the image unit after segmentation as the input data of geological disaster identification model, the efficiency of existing geological disaster visual interpretation can be effectively improved, data are provided and are supported for the prospecting of highway field operation and Hazard Risk Assessment.
Description
Technical field
The present invention relates to Highway Geological Disaster identification technology fields, and in particular to a kind of unmanned plane based on pre-training DCNN
Image Highway Geological Disaster recognition methods.
Background technique
Highway engineering because the distribution characteristics of its long ribbon shape make its inevitably need to pass through in construction it is different types of
Geomorphic unit is related to the landform and geological conditions of various complexity, the influence vulnerable to geological environment along the line.Especially mountain highway
After rain drop erosion or infiltration, the easily geological disasters such as induction avalanche, landslide, mud-rock flow.According to statistics, China is every year because of highway
It loses caused by geological disaster all up to tens yuan, brings about great losses to people's life safety and national wealth, geological disaster
Problem seriously threatens the normal operation of highway traffic infrastructure.Therefore, how Highway Geological Disaster is quickly and accurately obtained
Information has great importance for highway disaster reduction and prevention.
Traditional Highway Geological Disaster acquisition of information mostly uses the methods of gather material and the prospecting of field condition geological environment.
The problems such as that there are degrees of risk is high for field investigation, take time and effort, poor in timeliness.Unmanned aerial vehicle remote sensing is because flexibly lifting, low latitude fly for it
Row, the quickly advantages such as real-time, image resolution height, provide strong technology branch for Geological Hazards Investigation, monitoring and evaluation
It holds.However, interpretation efficiency is lower currently based on the Highway Geological Disaster identification of remote sensing image mostly based on artificial visual interpretation.
Summary of the invention
The object of the present invention is to provide a kind of unmanned plane image Highway Geological Disaster identification side based on pre-training DCNN
Method, realization is efficient, accurately interprets geological disaster.
To achieve the above object, the unmanned plane image Highway Geological Disaster based on pre-training DCNN designed by the present invention is known
Other method, comprising the following steps:
Step S1 obtains the unmanned aerial vehicle remote sensing images within the scope of road extension, is obtained after absolute orientation after pretreatment just
Projection picture；
Step S2 divides pretreated unmanned aerial vehicle remote sensing shadow using the mean shift algorithm for considering image texture feature
Picture；
Step S3 is applied to trained road geology using the unmanned aerial vehicle remote sensing images data of segmentation as input data
In disaster identification model, Highway Geological Disaster recognition result is obtained.
Preferably, the step S2 is further included steps of
Step S2.1 carries out average drifting processing to unmanned plane image, obtains the number for converging to the maximum point of probability density
According to；
Step S2.2 calculates the equivalent formulations LBP characteristic value of unmanned aerial vehicle remote sensing images after step S1 pretreatment；
Step S2.3, the image data based on step S2.1 and step S2.2 calculate space, color and the LBP line of image
Manage feature vector；
Step S2.4, the primary segmentation based on step S2.3 obtain final image as a result, carrying out region merging technique to image
Segmentation result.
Preferably, the step S2.1 is further included steps of
Step S2.1.1 calculates the mean shift vectors m (x) of pixel x according to following formula:
G (x)=- k'(x),
Wherein, h is bandwidth, and n is the quantity of pixel in bandwidth range, xiFor the ith pixel point in bandwidth range, d is
Space dimensionality, ‖ x ‖ indicate norm operation, and exp (x) is exponent arithmetic, and k (x) is the profile function of gaussian kernel function；
Step S2.1.2 sets allowable error ε, judgement | m (x)-x | the size with ε, if | m (x)-x | > ε assigns m (x)
To x, and return to step S2.1.1；If | m (x)-x | < ε terminates the iteration of pixel x；
Step S2.1.3 moves to next pixel x, repeats step S2.1.1~step S2.1.2, until on whole picture image
All pixels point.
Preferably, in the step S2.2, the calculating process of equivalent formulations LBP characteristic value is as follows:
The P circle shaped neighborhood regions that radius is R are set, with the central pixel point gray value g of centre of neighbourhood pixelcOn the basis of,
If central pixel point gray value gcGreatly, then location of pixels in contrast is labeled as 0；Conversely, being then labeled as 1, thus give birth to
At one group of binary sequence；
The transition times m in binary sequence by 0 to 1 or from 1 to 0 is counted, m≤2 are classified as equivalent formulations, by m
>'s 2 is classified as mixed mode, the transition times m of each pixel binary sequence of whole picture image statistics is traversed, by all equivalences
Mode is encoded to 0 according to+2 sequential encoding of 0~P (P-1), by mixed mode, and encoded radio is denoted as the equivalent formulations LBP of the pixel
Characteristic value.
Preferably, in the step S2.3, the equivalent formulations LBP texture calculated in step S2.2 is special
It levies vector to be added in the color feature vector of the pretreated image of step S2.1, installation space bandwidth parameter hs, color bandwidth
Parameter hrWith texture bandwidth parameter hl, space length will be met simultaneously less than hs, color characteristic distance is less than hrWith textural characteristics away from
From less than hlThe probability density maximum point of converging to cluster, obtain primary segmentation result.
Preferably, in the step S2.4, to image carry out region merging technique the step of it is as follows: setting face
Chromatic threshold value TrWith area threshold TsThe segmentation result of step S2.3 is based on Euclidean distance and carries out region merging technique, if adjacent two sub-district
The Euclidean distance of the spectrum characteristic parameter in domain is less than color threshold Tr, then it is merged into a region；If the picture in a certain region
First number is less than given area threshold Ts, then it is incorporated into the largest number of regions of pixel in adjacent area.
Preferably, in the step S3, trained Highway Geological Disaster identification model is according to as follows
Mode obtains:
Step S3.1 constructs DCNN linear ground object deleting madel；
Step S3.2 utilizes processed UCM remotely-sensed data collection training DCNN linear ground object deleting madel；
Step S3.3, DCNN linear ground object deleting madel parameter is moved to carried out on Highway Geological Disaster data set it is micro-
It adjusts, constructs Highway Geological Disaster identification model；
Step S3.4, the geological disaster remote sensing image data that the collection handled well is obtained are identified as Highway Geological Disaster
The input parameter of model, training Highway Geological Disaster identification model.
Preferably, in the step S3.1, DCNN linear ground object deleting madel includes 1 input layer, 4
A convolutional layer and pond layer, 2 full articulamentums and 1 output layer, output layer include linearly class and non-linear shape ground class classification
As a result；Each layer neuron is all made of ReLU type neuron in network: first layer convolutional layer window size is 5*5, convolution nuclear volume
It is 64；The second layer to the 4th layer of convolutional layer window size is 3*3, and convolution nuclear volume is 128, and the stride of convolutional layer is 1；
The window size of pond layer is 2*2, stride 2.
Preferably, in the step S3.2, by threadiness such as all roads, rivers in UCM data set
Ground class is set as negative sample, and it is uneven to handle positive negative sample using classification balance method as positive sample for the remotely-sensed data of remaining classification
Weighing apparatus, while data enhancing processing is carried out to image data collection, according to the positive and negative sample in 4:1 random division training set and test set
This.
Preferably, in the step S3.4, the specific step of the processing of geological disaster remote sensing image data
It is rapid as follows:
Step S3.4.1, collect geological disaster hotspots unmanned aerial vehicle remote sensing images data, to information on geological disasters into
Row visual interpretation, makes the vector file of geological disaster and non-geological disaster range, and batch cuts image data and according to unified
Format is stored, and obtains geological disaster positive sample and non-geological disaster negative example base, and all sample sizes are zoomed to system
One size；
Step S3.4.2 is marked geological disaster positive sample and non-geological disaster negative sample using marking tool；
Step S3.4.3 carries out Random Level, flip vertical, random scaling exptended sample collection to sample image；
Step S3.4.4 handles data nonbalance present in positive negative sample using classification balance method；
Step S3.4.5, upsets at random by step S3.4.3~step S3.4.4 treated data set sequence, according to
4:1 random division training set and test set guarantee that the positive and negative sample data ratio of training set and test set is 4:1.
Preferably, in the step S3, using the fixation window data of segmentation rear region as model
Input, fixed window are arranged as follows: the sum of all pixels of a certain region y after segmentation is denoted as Ny, each region is at least
The quantity for the pixel for including is Q, and the window sample size of region y is M × M (pixel):
If Ny≤ Q, the first similitude of zoning y and adjacent area, similarity measurement selecting index Distance conformability degree,
Similitude size is measured with two adjacent area centroid distances, region i is merged into the maximum adjacent area of similitude；Judge NyIt is
It is no to continue above-mentioned union operation if being unsatisfactory for condition greater than Q, until meeting Ny> Q then chooses with new combined region mass center
The image of M × M window size centered on pixel is denoted as whole region as mode input, the model recognition result of the window
Recognition result；
If Ny> Q then chooses the image of M × M window size using centered on regional quality imago element as mode input, is somebody's turn to do
The model recognition result of window is denoted as the recognition result of whole region.
The beneficial effects of the present invention are: the unmanned plane image Highway Geological Disaster identification of the invention based on pre-training DCNN
Method uses unmanned plane high resolution image, divides image based on the mean shift algorithm for considering textural characteristics, after segmentation
Input data of the image unit as geological disaster identification model exports disaster recognition result, can effectively improve existing geology calamity
The efficiency of evil visual interpretation provides data and supports, can effectively reduce field operation people for the prospecting of highway field operation and Hazard Risk Assessment
The personal safety risk of member, prevents and reduces natural disasters with great importance for highway engineering.
Detailed description of the invention
Fig. 1 is the unmanned plane image Highway Geological Disaster recognition methods based on pre-training DCNN of the preferred embodiment of the present invention
Flow chart.
Fig. 2 is the sub-process figure of the pretreatment unmanned aerial vehicle remote sensing images step in Fig. 1.
Fig. 3 is the sub-process figure of segmentation and treated image step in Fig. 1.
Fig. 4 is the sub-process figure for carrying out average drifting processing step in Fig. 3 to unmanned plane image.
Fig. 5 is the acquisition flow chart of the trained Highway Geological Disaster identification model in Fig. 1.
Fig. 6 is the structure chart of the Highway Geological Disaster identification model in Fig. 1.
Specific embodiment
The following further describes the present invention in detail with reference to the accompanying drawings and specific embodiments.
In recent years, with the rapid development of computer technology, deep learning has been widely used in computer vision, voice
The numerous areas such as identification and natural language processing.Convolutional neural networks (CNN, Convolutional Neural Networks)
It is most study in current deep learning, a most mature model of application, in high-resolution remote sensing image image classification, atural object
Extraction, scene Recognition etc. have been achieved for being applied successfully.However, CNN model training is needed using a large amount of marker samples,
Only in the case where training samples number is enough, network structure is more complicated, more excellent performance can be just shown；And
It lacks training in the case where sample, phenomena such as network is easy to appear over-fitting and falls into locally optimal solution.
The one kind of transfer learning as deep learning migration feature information can be used to solve target from existing data
Field lacks the problem concerning study of related training data, independent of the training data largely marked.Transfer learning is in source data
(large sample collection) field trains model in advance, and keeps pre-training aspect of model parameter constant, is applied to target data
Achieve the purpose that target identification after domain variability fine tuning model.In the case where lacking for Highway Geological Disaster training sample, introduce
Transfer learning mechanism can achieve the purpose of geological disaster efficient identification.
The problems such as efficiency is relatively low, and disaster training sample is few is interpreted for existing highway geological disaster, the present invention provides one kind
Unmanned plane image Highway Geological Disaster recognition methods based on pre-training DCNN will using unmanned plane high-resolution remote sensing image
The model feature information of pre-training is introduced into Highway Geological Disaster identification in transfer learning, finely tunes model realization Highway Geological Disaster
The rapidly extracting of information is effective way quick, that large area obtains disaster information, can reduce operating personnel's field condition and step on
The personal risk surveyed, while there is important guiding effect to highway engineering disaster reduction and prevention.
The unmanned plane image Highway Geological Disaster recognition methods based on pre-training DCNN that the invention discloses a kind of, incorporated by reference to
Refering to fig. 1, it the described method comprises the following steps:
Step S1 obtains the unmanned aerial vehicle remote sensing images within the scope of road extension, the DOM after absolute orientation is obtained after pretreatment
(Digital Orthophoto Map, digital orthoimage).
Step S2 divides pretreated unmanned aerial vehicle remote sensing shadow using the mean shift algorithm for considering image texture feature
Picture.
Step S3, using the unmanned aerial vehicle remote sensing images data of segmentation as input data, the good road geology calamity of application training
In evil identification model, road geology geological disaster recognition result is obtained.
Below with reference to attached drawing to the bright unmanned plane image Highway Geological Disaster identification side based on pre-training DCNN of we
Each step of method is described in detail.
In step sl, the unmanned aerial vehicle remote sensing images within the scope of road extension are obtained, after obtaining absolute orientation after pretreatment
Orthography.
It is measured using unmanned plane oblique photograph in road extension a certain range of pending geological disaster identification, utilizes nothing
Man-machine image processing software (for example, ENVI OneButton) carries out the even light of even color, empty three encryptions to image, generates absolute
Orthography after orientation cuts orthography, obtains the unmanned plane high-resolution number within the scope of pretreated road extension
According to, for subsequent geological disaster identification base image data are provided.Please refer to Fig. 2, specific steps are as follows:
Step S1.1 chooses road extension range, obtains highway according to the Highway Geological Disaster investigation that early period collects
The unmanned aerial vehicle remote sensing images of road domain range.For guarantee coverage geological disaster it is readily identified, preset unmanned plane during flying when
Ship's control is not less than 70%, and sidelapping degree is not less than 45%, and the ground resolution taken photo by plane reaches 0.2 meter.
Step S1.2 judges unmanned aerial vehicle remote sensing images with the presence or absence of loophole, defective data.Check unmanned plane image data matter
Amount is taken the photograph and is reseted to corresponding mend is carried out there are the data of loophole, defect, is required with meeting orthography production.
Step S1.3 carries out photogrammetric distortion correction, even color dodging to image.Since awing posture is unstable for unmanned plane
It is fixed, cause the image obtained there are distortion phenomenon, needs to carry out photogrammetric distortion correction to image.Furthermore unmanned plane obtain image because
Exposure time interval, weather difference will lead to image data there are differences such as color, illumination, need to carry out image even color, even
Light processing.
Step S1.4 carries out analytical aerial triangulation (sky three encrypts) and block adjustment to image, after generating correction
Orthography data.Three encryption of sky and regional network are carried out to image using ENVI OneButton unmanned plane image processing software
Adjustment, the orthography data after generating correction.
Step S1.5 cuts orthography, obtains the unmanned plane high-resolution number within the scope of pretreated road extension
According to.
In step s 2, using the pretreated unmanned plane shadow of mean shift algorithm segmentation for considering image texture feature
Picture.
Existing convolutional neural networks Remote Image Classification is it needs to be determined that classification belonging to each pixel, this processing
Method efficiency is lower.Identical or phase is shown under the conditions ofs identical landform, illumination etc. according to the similar atural object in remote sensing image
As spectral information and spatial information feature, consider the texture feature information of image, using consider textural characteristics average drifting
Pretreated orthography data in algorithm segmentation step S1.5, the input data as subsequent geological disaster identification model.
The essence of mean shift algorithm is a kind of adaptive fast-rise approach, and the mean value by calculating each pixel is floated
Vector is moved, the difference of average drifting vector and sampled point is compared, circulation asks difference until being less than given threshold value, that is, reaches convergence
To the maximum point of probability density.Fig. 3 is please referred to, steps are as follows for the mean shift segmentation of remote sensing image:
Step S2.1 carries out average drifting processing to unmanned plane image, obtains the number for converging to the maximum point of probability density
According to.Fig. 4 is please referred to, the process of average drifting processing is as follows:
Step S2.1.1 calculates the mean shift vectors m (x) of a certain pixel x in image according to following formula:
G (x)=- k'(x) (2)
Wherein, h is bandwidth, and n is the quantity of pixel in bandwidth range, xiFor the ith pixel point in bandwidth range, d is
Space dimensionality；‖ x ‖ indicates norm operation, and exp (x) is exponent arithmetic, and k (x) is the profile function of gaussian kernel function.
Initial pixel point x is given as the 1st pixel in the image upper left corner to be split in step S1.5.
Step S2.1.2 is set allowable error ε (such as value as 0.01), judgement | m (x)-x | the size with ε.
If | m (x)-x | m (x) is assigned to x, and returns to step S2.1.1 by > ε；If | m (x)-x | < ε terminates the pixel
The iteration of x.
Step S2.1.3 moves to next pixel x, repeats step S2.1.1~step S2.1.2, until on whole picture image
All pixels point, complete filtering processing to entire image, obtain the data for converging to the maximum point of probability density.
Step S2.2 calculates the equivalent formulations LBP characteristic value of unmanned aerial vehicle remote sensing images after step S1 pretreatment.
When traditional mean shift algorithm carries out Image Segmentation, only considers spatial position and color characteristic information, do not examine
Consider the texture feature information of image.Uniform-LBP (Local Binary Pattern, local binary patterns) operator can be retouched
The image texture information in different scale is stated, and the monotone variation of any gray scale is able to maintain and is basically unchanged, while can be had
Effect avoids data volume excessive, and most key messages of image can be retained while carrying out dimensionality reduction to data.Therefore, of the invention
Uniform-LBP operator is introduced into Image Segmentation, using the textural characteristics of Uniform-LBP operator representation image, obtain etc.
Valence mode LBP characteristic value, calculating process are as follows:
The P circle shaped neighborhood regions that radius is R are set, with the central pixel point gray value g of centre of neighbourhood pixelcOn the basis of,
By the gray value g of itself and other pixels of P vertex neighborhoodpCompare.If central pixel point gray value gcIt greatly, then will be right with it
The location of pixels of ratio is labeled as 0；Conversely, being then labeled as 1, one group of binary sequence is thus generated.
The transition times m in binary sequence by 0 to 1 or from 1 to 0 is counted, m≤2 are classified as equivalent formulations, by m
>'s 2 is classified as mixed mode.Whole picture image is traversed, the transition times m of each pixel binary sequence is counted.By all equivalences
Mode encodes in sequence, and encoded radio range is 0~P (P-1)+2, obtains [0,1,2,3,4,5,6 ... ..., P (P-1)+2]；It is mixed
Syntype is encoded to 0, and encoded radio is denoted as the equivalent formulations LBP characteristic value of the pixel.
Step S2.3, the image data based on step S2.1 and step S2.2 calculate space, color and the LBP line of image
Manage feature vector.
The pretreated image of step S2.1 is added in the equivalent formulations LBP texture feature vector calculated in step S2.2
In color feature vector, installation space bandwidth parameter hs, color bandwidth parameter hrWith texture bandwidth parameter hl, sky will be met simultaneously
Between distance be less than hs, color characteristic distance is less than hrIt is less than h with textural characteristics distancelConverge to probability density maximum point
It is clustered, obtains primary segmentation result.
Step S2.4, the primary segmentation based on step S2.3 obtain final image as a result, carrying out region merging technique to image
Segmentation result.
Combined process is as follows: setting color threshold TrWith area threshold TsThe segmentation result of step S2.3 is based on European
Distance (euclidean metric) carries out region merging technique.If the Euclidean distance of the spectrum characteristic parameter of adjacent two subregion is less than color threshold
Value Tr, then it is merged into a region；If the pixel number in a certain region is less than given area threshold Ts, then closed
And the largest number of regions of pixel into adjacent area, obtain final segmentation result.
In step s3, the unmanned plane image divided in step S2 is inputted into trained Highway Geological Disaster and identifies mould
Type exports Highway Geological Disaster recognition result.
The unmanned aerial vehicle remote sensing images data of the mean shift algorithm segmentation of textural characteristics will be considered as defeated in step S2
Enter data, be applied in trained Highway Geological Disaster identification model, exports geological disaster recognition result.
Since the area size after Image Segmentation is different, area zoom to same size directly can be lost into image information,
Therefore, using the fixation window data of segmentation rear region as mode input, fixed window is provided that a certain area after segmentation
The sum of all pixels of domain y is denoted as Ny, the quantity for the pixel that each region includes at least is Q, and the window sample size of region y is M
× M (pixel):
1. if Ny≤ Q, the first similitude of zoning y and adjacent area, similarity measurement selecting index is apart from similar
Degree measures similitude size with two adjacent area centroid distances, region i is merged into the maximum adjacent area of similitude；Judgement
NyWhether it is greater than Q and continues above-mentioned union operation if being unsatisfactory for condition, until meets Ny> Q then chooses with new combined region matter
For the image of M × M window size centered on imago element as mode input, the model recognition result of the window is denoted as whole region
Recognition result.
2. if Ny> Q, then choose the image of M × M window size using centered on regional quality imago element as mode input,
The model recognition result of the window is denoted as the recognition result of whole region.
After determining fixed window, using the image data of window as mode input, geology is carried out to the image data of window
Disaster differentiates whether the probabilistic determination according to result, that is, geological disaster of model training and non-geological disaster classification is geology calamity
Evil obtains the geological disaster identification output result of highway remote sensing image.
In above-mentioned steps S3, referring to Fig. 5, trained Highway Geological Disaster identification model obtains as follows
It arrives:
Step S3.1 constructs DCNN linear ground object deleting madel.Build the depth convolutional neural networks DCNN of pre-training
(Deep Convolutional Neural Network) linear ground object deleting madel frame, including 1 input layer, 4 convolution
Layer (convolutional layer+pond layer), 2 full articulamentums, 1 output layer, output layer include two category results, and classification 1 is linearly
Class, for non-linear shape the class of classification 2.Wherein, in network each layer neuron be all made of ReLU (Rectified Linear Units,
Correct linear unit) type neuron: first layer convolutional layer window size is 5*5, and convolution nuclear volume is 64；The second layer is to the 4th
Layer convolutional layer window size is 3*3, and convolution nuclear volume is 128, and the stride of convolutional layer is 1；The window size of pond layer is equal
For 2*2, stride 2, as shown in Figure 6.Output layer includes two category results, and classification 1 is that linearly class, classification 2 are non-thread
Shape ground class.
Step S3.2 utilizes processed UCM remotely-sensed data collection training DCNN linear ground object deleting madel.Utilize UC
Merced Land-Use Dataset (hereinafter referred to as UCM data set) constructs the DCNN linear ground object deleting madel of pre-training,
For the output result of the model there are two classification, classification 1 is linearly class (road, river etc.) to be rejected, and classification 2 is non-linear shape
Ground class.
Traditional deep learning needs to mark a large amount of training data to target domain, can be difficult if training sample is very few
To construct deep learning model, cause correlative study that can not carry out with application.Transfer learning is one kind of deep learning, Ke Yicong
Migration feature information is for the learning tasks in target domain in existing mass data collection (source domain), independent of a large amount of
Training sample support, and can achieve the purpose that geological disaster efficient identification.Transfer learning is at source data (large sample collection)
Field trains model in advance, and keeps pre-training aspect of model parameter constant, is applied to the fine tuning of target data domain variability
Achieve the purpose that target identification after model.
Lack situation for Highway Geological Disaster sample, transfer learning method is introduced Highway Geological Disaster and identifies mould by the present invention
It is the interference for reducing non-targeted atural object as far as possible in type, the depth convolutional neural networks DCNN of building pre-training first is linearly
Object deleting madel, training data use UCM data set, and UCM data set includes 21 class remote sensing scene images, the output knot of model altogether
For fruit there are two classification, classification 1 is linearly class (road, river etc.) to be rejected, for non-linear shape the class of classification 2.
By all roads, river in UCM data set etc., linearly class is set as negative sample, and the remotely-sensed data of remaining classification is made
For positive sample, the problem of positive and negative sample imbalance is handled using classification balance method, while data increasing is carried out to image data collection
Strength reason, using the data handled well as the training sample of pre-training DCNN linearly class deleting madel, according to 4:1 random division
Positive negative sample in training set and test set.
Above-mentioned " classification balance method " can be used Haikang Wei Shi research institute Shicai Yang etc. and appoint in ILSVRC scene classification
The balance method of " the classification recombination " that proposes in business, the steps include: firstly, being ranked up according to classification sequence to original sample；
Later, it calculates the number of samples of each classification and records the number of samples of sample at most that class；According to this most sample number
The list that a random alignment is generated to every class sample, is then taken with sample number of the random number in this list to respective classification
It is remaining, obtain corresponding index value；Then, image is extracted from such image according to index, generates such image random column
Table；The random list of all classes is connected together and upsets order at random, final image list can be obtained, it can be found that finally
Every class number of samples is impartial in list.
Utilize sample data, training depth convolutional neural networks DCNN linear ground object deleting madel.
Step S3.3 constructs Highway Geological Disaster identification model.Fig. 6 is please referred to, keeps pre-training DCNN linearly
The characteristic parameter of object deleting madel is constant, is moved to and is finely adjusted on Highway Geological Disaster data set, connect 1 it is new defeated
Layer out, output layer include two new classifications, and classification 1 is geological disaster as a result, classification 2 is non-geological disaster as a result, building public affairs
Road geological disaster identification model.
Based on migration results, the feature vector that feature migration phase is exported is as the training input of SVM classifier, building
Highway Geological Disaster identification model.
Step S3.4, training Highway Geological Disaster identification model.Using the geological disaster remote sensing image data handled well as
The input parameter of Highway Geological Disaster identification model starts training pattern.
Prepare the training and the positive and negative sample data of test of Highway Geological Disaster, and marker samples classification information, as geology
The training data of disaster identification model.Specific step is as follows for the processing of geological disaster remote sensing image data:
Step S3.4.1, collect geological disaster hotspots unmanned aerial vehicle remote sensing images data, to information on geological disasters into
Row visual interpretation makes geological disaster (avalanche, landslide, mud-rock flow disaster) using geographical handling implement (such as ArcGIS software)
With the vector file of non-geological disaster range, batch cuts image data, and is stored according to unified format, obtains geology calamity
Evil positive sample and non-geological disaster negative example base, and all sample sizes are zoomed into unified size.
Step S3.4.2, convolutional neural networks training pattern need sample class to be marked classification, utilize label work
Geological disaster positive sample is marked in tool (such as LabelImg), is labeled as 1, is labeled as 0 to non-geological disaster negative sample.
Step S3.4.3 carries out Random Level, flip vertical, random scaling exptended sample collection to sample image.Due to number
It is easy to cause model to show over-fitting in training according to sample size is very few, therefore, to avoid model over-fitting to a certain extent,
A data enhancing processing is carried out to the markd positive and negative sample data of institute, Random Level, flip vertical, random is carried out to sample image
Scale exptended sample collection.
Step S3.4.4, since non-geological disaster sample size is more than geological disaster sample size, using classification balance side
Method (balance method that Haikang prestige regards graduate " classification recombination ") solves the problems, such as data nonbalance present in positive negative sample, protects
The probability for demonstrate,proving each classification participation training is more balanced.
Step S3.4.5, upsets at random by step S3.4.3~step S3.4.4 treated data set sequence, according to
4:1 random division training set and test set guarantee that the positive and negative sample data ratio of training set and test set is 4:1.
Training data of the disaster data handled well using above-mentioned steps as Highway Geological Disaster identification model, using few
Training Highway Geological Disaster identification model is finely tuned in the geological disaster and non-geological disaster data for measuring tape label information.
Unmanned plane image Highway Geological Disaster recognition methods based on pre-training DCNN of the invention is intended to transfer learning
Mechanism is introduced into Highway Geological Disaster identification, and by constructing pre-training DCNN model, the characteristic parameter of pre-training model is migrated
After being finely adjusted training on to Highway Geological Disaster sample database, achieve the purpose that Highway Geological Disaster quickly identifies.With Yunnan Province
For the horse highway of river in Shangdong Province, complex geologic conditions along the horse highway of river in Shangdong Province pass through a plurality of Earthquake Fault Zone, by geological disaster shadow
Sound is larger, identifies through the invention to highroad geological disaster, can effectively reduce highway field operation early period survey amount,
Field personnel operating risk is reduced, prospecting efficiency is improved.
Compared with prior art, the unmanned plane image Highway Geological Disaster identification side of the invention based on pre-training DCNN
Method has the advantage that
(1) transfer learning mechanism is introduced into Highway Geological Disaster identification, using unmanned plane high resolution image, based on examining
The mean shift algorithm for considering textural characteristics divides image, using the image unit after segmentation as the input of geological disaster identification model
Data export disaster recognition result, can effectively improve the efficiency of existing geological disaster visual interpretation；
(2) to can solve existing information on geological disasters training sample by transfer learning few, it is difficult to build depth nerve net
Network model problem is realized that Highway Geological Disaster progress is quick, macroscopic view, is accurately identified, for the prospecting of highway field operation and disasters danger
Evaluation provides data and supports, can effectively reduce the personal safety risk of outdoor workers, and preventing and reducing natural disasters for highway engineering has weight
Want directive significance.
The embodiments described above only express several embodiments of the present invention, and the description thereof is more specific and detailed, but simultaneously
Limitations on the scope of the patent of the present invention therefore cannot be interpreted as.It should be pointed out that for those of ordinary skill in the art
For, without departing from the inventive concept of the premise, various modifications and improvements can be made, these belong to guarantor of the invention
Protect range.Therefore, the scope of protection of the patent of the invention shall be subject to the appended claims.
Claims (11)
1. a kind of unmanned plane image Highway Geological Disaster recognition methods based on pre-training DCNN, comprising the following steps:
Step S1 obtains the unmanned aerial vehicle remote sensing images within the scope of road extension, the orthogonal projection after absolute orientation is obtained after pretreatment
Picture；
Step S2 divides pretreated unmanned aerial vehicle remote sensing images using the mean shift algorithm for considering image texture feature；
Step S3 is applied to trained Highway Geological Disaster using the unmanned aerial vehicle remote sensing images data of segmentation as input data
In identification model, Highway Geological Disaster recognition result is obtained.
2. the unmanned plane image Highway Geological Disaster recognition methods according to claim 1 based on pre-training DCNN, special
Sign is that the step S2 is further included steps of
Step S2.1 carries out average drifting processing to unmanned plane image, obtains the data for converging to the maximum point of probability density；
Step S2.2 calculates the equivalent formulations LBP characteristic value of unmanned aerial vehicle remote sensing images after step S1 pretreatment；
Step S2.3, the image data based on step S2.1 and step S2.2, space, color and the LBP texture for calculating image are special
Levy vector；
Step S2.4, the primary segmentation based on step S2.3 obtain final Image Segmentation as a result, carrying out region merging technique to image
As a result.
3. the unmanned plane image Highway Geological Disaster recognition methods according to claim 2 based on pre-training DCNN, special
Sign is that the step S2.1 is further included steps of
Step S2.1.1 calculates the mean shift vectors m (x) of pixel x according to following formula:
G (x)=- k ' (x),
Wherein, h is bandwidth, and n is the quantity of pixel in bandwidth range, xiFor the ith pixel point in bandwidth range, d is space dimension
Number, ‖ x ‖ indicate norm operation, and exp (x) is exponent arithmetic, and k (x) is the profile function of gaussian kernel function；
Step S2.1.2 sets allowable error ε, judgement | m (x)-x | the size with ε, if | m (x)-x | m (x) is assigned to x by > ε,
And return to step S2.1.1；If | m (x)-x | < ε terminates the iteration of pixel x；
Step S2.1.3 moves to next pixel x, repeats step S2.1.1~step S2.1.2, until the institute on whole picture image
There is pixel.
4. the unmanned plane image Highway Geological Disaster recognition methods according to claim 2 based on pre-training DCNN, special
Sign is, in the step S2.2, the calculating process of equivalent formulations LBP characteristic value is as follows:
The P circle shaped neighborhood regions that radius is R are set, with the central pixel point gray value g of centre of neighbourhood pixelcOn the basis of, if in
Heart pixel gray value gcGreatly, then location of pixels in contrast is labeled as 0；Conversely, being then labeled as 1, one group is thus generated
Binary sequence；
The transition times m in binary sequence by 0 to 1 or from 1 to 0 is counted, m≤2 are classified as equivalent formulations, by m >'s 2
It is classified as mixed mode, the transition times m of each pixel binary sequence of whole picture image statistics is traversed, all equivalent formulations is pressed
According to+2 sequential encoding of 0~P (P-1), mixed mode is encoded to 0, encoded radio is denoted as the equivalent formulations LBP feature of the pixel
Value.
5. the unmanned plane image Highway Geological Disaster recognition methods according to claim 2 based on pre-training DCNN, special
Sign is: in the step S2.3, step S2.1 is added in the equivalent formulations LBP texture feature vector calculated in step S2.2
In the color feature vector of pretreated image, installation space bandwidth parameter hs, color bandwidth parameter hrWith texture bandwidth parameter
hl, space length will be met simultaneously less than hs, color characteristic distance is less than hrIt is less than h with textural characteristics distancelConverge to probability
The maximum point of density is clustered, and primary segmentation result is obtained.
6. the unmanned plane image Highway Geological Disaster recognition methods according to claim 2 based on pre-training DCNN, special
Sign is: in the step S2.4, to image carry out region merging technique the step of it is as follows: setting color threshold TrAnd area threshold
TsThe segmentation result of step S2.3 is based on Euclidean distance and carries out region merging technique, if the Europe of the spectrum characteristic parameter of adjacent two subregion
Formula distance is less than color threshold Tr, then it is merged into a region；If the pixel number in a certain region is less than given face
Product threshold value Ts, then it is incorporated into the largest number of regions of pixel in adjacent area.
7. the unmanned plane image Highway Geological Disaster according to any one of claim 2 to 6 based on pre-training DCNN is known
Other method, which is characterized in that in the step S3, trained Highway Geological Disaster identification model obtains as follows
It arrives:
Step S3.1 constructs DCNN linear ground object deleting madel；
Step S3.2 utilizes processed UCM remotely-sensed data collection training DCNN linear ground object deleting madel；
DCNN linear ground object deleting madel parameter is moved to and is finely adjusted on Highway Geological Disaster data set by step S3.3, structure
Build Highway Geological Disaster identification model；
Step S3.4, the geological disaster remote sensing image data that the collection handled well is obtained is as Highway Geological Disaster identification model
Input parameter, training Highway Geological Disaster identification model.
8. the unmanned plane image Highway Geological Disaster recognition methods according to claim 7 based on pre-training DCNN, special
Sign is: in the step S3.1, DCNN linear ground object deleting madel includes 1 input layer, 4 convolutional layers and pond layer, 2
A full articulamentum and 1 output layer, output layer include linearly class and non-linear shape ground class category result；Each layer nerve in network
Member is all made of ReLU type neuron: first layer convolutional layer window size is 5*5, and convolution nuclear volume is 64；The second layer is to the 4th
Layer convolutional layer window size is 3*3, and convolution nuclear volume is 128, and the stride of convolutional layer is 1；The window size of pond layer is equal
For 2*2, stride 2.
9. the unmanned plane image Highway Geological Disaster recognition methods according to claim 7 based on pre-training DCNN, special
Sign is: in the step S3.2, by all roads, river in UCM data set etc., linearly class is set as negative sample, remaining
The remotely-sensed data of classification handles positive and negative sample imbalance as positive sample, using classification balance method, while to image data collection
Data enhancing processing is carried out, according to the positive negative sample in 4:1 random division training set and test set.
10. the unmanned plane image Highway Geological Disaster recognition methods according to claim 7 based on pre-training DCNN, special
Sign is: in the step S3.4, specific step is as follows for the processing of geological disaster remote sensing image data:
Step S3.4.1 collects the unmanned aerial vehicle remote sensing images data of geological disaster hotspots, carries out mesh to information on geological disasters
Depending on interpretation, the vector file of geological disaster and non-geological disaster range is made, batch cuts image data and according to unified format
It is stored, obtains geological disaster positive sample and non-geological disaster negative example base, and all sample sizes are zoomed into unification greatly
It is small；
Step S3.4.2 is marked geological disaster positive sample and non-geological disaster negative sample using marking tool；
Step S3.4.3 carries out Random Level, flip vertical, random scaling exptended sample collection to sample image；
Step S3.4.4 handles the data nonbalance of positive negative sample using classification balance method；
Step S3.4.5, upsets at random by step S3.4.3~step S3.4.4 treated data set sequence, according to 4:1 with
Machine divides training set and test set, guarantees that the positive and negative sample data ratio of training set and test set is 4:1.
11. the unmanned plane image Highway Geological Disaster recognition methods according to claim 1 based on pre-training DCNN, special
Sign is: in the step S3, using the fixation window data of segmentation rear region as mode input, fixed window is according to such as
Under type setting: the sum of all pixels of a certain region y after segmentation is denoted as Ny, the quantity for the pixel that each region includes at least is
The window sample size of Q, region y are M × M (pixel):
If Ny≤ Q, the first similitude of zoning y and adjacent area, similarity measurement selecting index Distance conformability degree, with two
Adjacent area centroid distance measures similitude size, and region i is merged into the maximum adjacent area of similitude；Judge NyIt is whether big
In Q, if being unsatisfactory for condition, continue above-mentioned union operation, until meeting Ny> Q then chooses with new combined region mass center pixel
Centered on M × M window size image as mode input, the model recognition result of the window is denoted as the identification of whole region
As a result；
If Ny> Q then chooses the image of M × M window size using centered on regional quality imago element as mode input, the window
Model recognition result be denoted as the recognition result of whole region.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910204723.6A CN110008854B (en) | 2019-03-18 | 2019-03-18 | Unmanned aerial vehicle image highway geological disaster identification method based on pre-training DCNN |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910204723.6A CN110008854B (en) | 2019-03-18 | 2019-03-18 | Unmanned aerial vehicle image highway geological disaster identification method based on pre-training DCNN |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110008854A true CN110008854A (en) | 2019-07-12 |
| CN110008854B CN110008854B (en) | 2021-04-30 |
Family
ID=67167296
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910204723.6A Active CN110008854B (en) | 2019-03-18 | 2019-03-18 | Unmanned aerial vehicle image highway geological disaster identification method based on pre-training DCNN |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110008854B (en) |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110619283A (en) * | 2019-08-26 | 2019-12-27 | 海南撰云空间信息技术有限公司 | Automatic extraction method for unmanned aerial vehicle ortho-image road |
| CN111460965A (en) * | 2020-03-27 | 2020-07-28 | 珠海欧比特宇航科技股份有限公司 | Waterlogging area intelligent analysis system and method based on unmanned aerial vehicle image |
| CN111860236A (en) * | 2020-07-06 | 2020-10-30 | 中国科学院空天信息创新研究院 | Small sample remote sensing target detection method and system based on transfer learning |
| CN111931738A (en) * | 2020-09-28 | 2020-11-13 | 南京邮电大学 | Neural network model pre-training method and device for remote sensing image |
| CN112326686A (en) * | 2020-11-02 | 2021-02-05 | 坝道工程医院(平舆) | Unmanned aerial vehicle intelligent cruise pavement disease detection method, unmanned aerial vehicle and detection system |
| CN112344913A (en) * | 2020-11-09 | 2021-02-09 | 中国公路工程咨询集团有限公司 | Regional risk coefficient evaluation method by utilizing oblique photography image of unmanned aerial vehicle |
| CN112634169A (en) * | 2020-12-30 | 2021-04-09 | 成都星时代宇航科技有限公司 | Color homogenizing method and device for remote sensing image |
| CN112966548A (en) * | 2021-01-18 | 2021-06-15 | 中国农业科学院农业信息研究所 | Soybean plot identification method and system |
| CN113191177A (en) * | 2020-11-20 | 2021-07-30 | 北京林业大学 | Pine wood nematode remote sensing image sample labeling method for deep learning |
| CN114639008A (en) * | 2020-12-01 | 2022-06-17 | 中移(成都)信息通信科技有限公司 | Method, device and equipment for extracting remote sensing image building and computer storage medium |
| CN114821376A (en) * | 2022-06-27 | 2022-07-29 | 中咨数据有限公司 | Unmanned aerial vehicle image geological disaster automatic extraction method based on deep learning |
| GB2621645A (en) * | 2022-06-27 | 2024-02-21 | Checc Data Co Ltd | Deep learning based method for automatic geological disaster extraction from unmanned aerial vehicle image |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101706950A (en) * | 2009-07-13 | 2010-05-12 | 中国科学院遥感应用研究所 | High-performance implementation method for multi-scale segmentation of remote sensing images |
| CN106169086A (en) * | 2016-07-21 | 2016-11-30 | 武汉大学 | High-resolution optical image damage method for extracting roads under navigation data auxiliary |
| CN106355248A (en) * | 2016-08-26 | 2017-01-25 | 深圳先进技术研究院 | Deep convolution neural network training method and device |
| CN107194349A (en) * | 2017-05-19 | 2017-09-22 | 湖北工业大学 | High-resolution PolSAR images city terrain classification method based on deep learning |
| CN107578003A (en) * | 2017-08-29 | 2018-01-12 | 中国科学院遥感与数字地球研究所 | A kind of remote sensing images transfer learning method based on GEOGRAPHICAL INDICATION image |
-
2019
- 2019-03-18 CN CN201910204723.6A patent/CN110008854B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101706950A (en) * | 2009-07-13 | 2010-05-12 | 中国科学院遥感应用研究所 | High-performance implementation method for multi-scale segmentation of remote sensing images |
| CN106169086A (en) * | 2016-07-21 | 2016-11-30 | 武汉大学 | High-resolution optical image damage method for extracting roads under navigation data auxiliary |
| CN106355248A (en) * | 2016-08-26 | 2017-01-25 | 深圳先进技术研究院 | Deep convolution neural network training method and device |
| CN107194349A (en) * | 2017-05-19 | 2017-09-22 | 湖北工业大学 | High-resolution PolSAR images city terrain classification method based on deep learning |
| CN107578003A (en) * | 2017-08-29 | 2018-01-12 | 中国科学院遥感与数字地球研究所 | A kind of remote sensing images transfer learning method based on GEOGRAPHICAL INDICATION image |
Non-Patent Citations (3)
| Title |
|---|
| 张伟: "遥感震害信息智能提取优化方法研究", 《中国优秀硕士学位论文全文数据库基础科学辑》 * |
| 李婷婷等: "基于改进的均值漂移算法的高分辨率遥感影像分割方法研究", 《贵州大学学报( 自然科学版)》 * |
| 沙爱民等: "基于卷积神经网络的路表病害识别与测量", 《中国公路学报》 * |
Cited By (21)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110619283A (en) * | 2019-08-26 | 2019-12-27 | 海南撰云空间信息技术有限公司 | Automatic extraction method for unmanned aerial vehicle ortho-image road |
| CN110619283B (en) * | 2019-08-26 | 2023-01-10 | 海南撰云空间信息技术有限公司 | Automatic extraction method for unmanned aerial vehicle ortho-image road |
| CN111460965A (en) * | 2020-03-27 | 2020-07-28 | 珠海欧比特宇航科技股份有限公司 | Waterlogging area intelligent analysis system and method based on unmanned aerial vehicle image |
| CN111460965B (en) * | 2020-03-27 | 2024-05-31 | 珠海欧比特宇航科技股份有限公司 | Waterlogging area intelligent analysis system and method based on unmanned aerial vehicle image |
| CN111860236B (en) * | 2020-07-06 | 2021-12-03 | 中国科学院空天信息创新研究院 | Small sample remote sensing target detection method and system based on transfer learning |
| CN111860236A (en) * | 2020-07-06 | 2020-10-30 | 中国科学院空天信息创新研究院 | Small sample remote sensing target detection method and system based on transfer learning |
| CN111931738A (en) * | 2020-09-28 | 2020-11-13 | 南京邮电大学 | Neural network model pre-training method and device for remote sensing image |
| CN111931738B (en) * | 2020-09-28 | 2021-01-26 | 南京邮电大学 | Neural network model pre-training method and device for remote sensing image |
| CN112326686A (en) * | 2020-11-02 | 2021-02-05 | 坝道工程医院(平舆) | Unmanned aerial vehicle intelligent cruise pavement disease detection method, unmanned aerial vehicle and detection system |
| CN112326686B (en) * | 2020-11-02 | 2024-02-02 | 坝道工程医院(平舆) | Unmanned aerial vehicle intelligent cruising pavement disease detection method, unmanned aerial vehicle and detection system |
| CN112344913A (en) * | 2020-11-09 | 2021-02-09 | 中国公路工程咨询集团有限公司 | Regional risk coefficient evaluation method by utilizing oblique photography image of unmanned aerial vehicle |
| CN113191177A (en) * | 2020-11-20 | 2021-07-30 | 北京林业大学 | Pine wood nematode remote sensing image sample labeling method for deep learning |
| CN114639008A (en) * | 2020-12-01 | 2022-06-17 | 中移(成都)信息通信科技有限公司 | Method, device and equipment for extracting remote sensing image building and computer storage medium |
| CN112634169A (en) * | 2020-12-30 | 2021-04-09 | 成都星时代宇航科技有限公司 | Color homogenizing method and device for remote sensing image |
| CN112634169B (en) * | 2020-12-30 | 2024-02-27 | 成都星时代宇航科技有限公司 | Remote sensing image color homogenizing method and device |
| CN112966548A (en) * | 2021-01-18 | 2021-06-15 | 中国农业科学院农业信息研究所 | Soybean plot identification method and system |
| CN112966548B (en) * | 2021-01-18 | 2023-10-27 | 中国农业科学院农业信息研究所 | Soybean plot identification method and system |
| CN114821376A (en) * | 2022-06-27 | 2022-07-29 | 中咨数据有限公司 | Unmanned aerial vehicle image geological disaster automatic extraction method based on deep learning |
| CN114821376B (en) * | 2022-06-27 | 2022-09-20 | 中咨数据有限公司 | Unmanned aerial vehicle image geological disaster automatic extraction method based on deep learning |
| WO2024000927A1 (en) * | 2022-06-27 | 2024-01-04 | 中咨数据有限公司 | Deep learning based method for automatic geological disaster extraction from unmanned aerial vehicle image |
| GB2621645A (en) * | 2022-06-27 | 2024-02-21 | Checc Data Co Ltd | Deep learning based method for automatic geological disaster extraction from unmanned aerial vehicle image |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110008854B (en) | 2021-04-30 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110008854A (en) | Unmanned plane image Highway Geological Disaster recognition methods based on pre-training DCNN | |
| CN108710875B (en) | A kind of take photo by plane road vehicle method of counting and device based on deep learning | |
| CN107092020B (en) | Merge the surface evenness monitoring method of unmanned plane LiDAR and high score image | |
| CN112287807B (en) | Remote sensing image road extraction method based on multi-branch pyramid neural network | |
| Sun et al. | Modeling urban land use change and urban sprawl: Calgary, Alberta, Canada | |
| CN109697475A (en) | A kind of muskeg information analysis method, remote sensing monitoring component and monitoring method | |
| CN109241846A (en) | Change in time and space estimating and measuring method, device and the storage medium of remote sensing image | |
| CN112668375B (en) | Tourist distribution analysis system and method in scenic spot | |
| CN109299673A (en) | The green degree spatial extraction method of group of cities and medium | |
| Yuan et al. | Fine-grained classification of urban functional zones and landscape pattern analysis using hyperspectral satellite imagery: A case study of Wuhan | |
| CN113610070A (en) | Landslide disaster identification method based on multi-source data fusion | |
| CN115761513A (en) | Intelligent remote sensing identification method for mountain large landslide based on semi-supervised deep learning | |
| CN114898089B (en) | Functional area extraction and classification method fusing high-resolution images and POI data | |
| CN108154089B (en) | Size-adaptive-based crowd counting method for head detection and density map | |
| Vaidya et al. | Classifying heterogeneous urban form into local climate zones using supervised learning and greedy clustering incorporating Landsat dataset | |
| CN114067245A (en) | Method and system for identifying hidden danger of external environment of railway | |
| CN116883854A (en) | Detection model and method for identifying hydraulic building in satellite remote sensing image | |
| Yadav et al. | Automatic urban road extraction from high resolution satellite data using object based ımage analysis: A fuzzy classification approach | |
| Jiao et al. | A Novel Data Augmentation Method to Enhance the Training Dataset for Road Extraction from Historical Maps | |
| CN110363101A (en) | A kind of flowers recognition methods based on CNN Fusion Features frame | |
| Huang et al. | Simultaneous update of high-resolution land-cover mapping attempt: Wuhan and the surrounding satellite cities cartography using L2HNet | |
| Song | The Fine Design Strategy of Urban Streets Using Deep Learning with the Assistance of the Internet of Things | |
| Karila et al. | Automatic labelling for semantic segmentation of VHR satellite images: Application of airborne laser scanner data and object-based image analysis | |
| Li et al. | “3S” Technologies and Application for Dynamic Monitoring Soil and Water Loss in the Yangtze River Basin, China | |
| Djenaliev et al. | Extraction of built-up areas from Landsatimagery using the object-oriented classificationmethod |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |