CN115409842B - Remote sensing image unsupervised change detection method based on convolution self-encoder-decoder - Google Patents
Remote sensing image unsupervised change detection method based on convolution self-encoder-decoder Download PDFInfo
- Publication number
- CN115409842B CN115409842B CN202211358970.XA CN202211358970A CN115409842B CN 115409842 B CN115409842 B CN 115409842B CN 202211358970 A CN202211358970 A CN 202211358970A CN 115409842 B CN115409842 B CN 115409842B
- Authority
- CN
- China
- Prior art keywords
- remote sensing
- change detection
- periods
- loss
- 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.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/74—Image or video pattern matching; Proximity measures in feature spaces
- G06V10/761—Proximity, similarity or dissimilarity measures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
- G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10024—Color image
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10032—Satellite or aerial image; Remote sensing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Computation (AREA)
- General Physics & Mathematics (AREA)
- Physics & Mathematics (AREA)
- Computing Systems (AREA)
- General Health & Medical Sciences (AREA)
- Medical Informatics (AREA)
- Software Systems (AREA)
- Databases & Information Systems (AREA)
- Artificial Intelligence (AREA)
- Health & Medical Sciences (AREA)
- Multimedia (AREA)
- Quality & Reliability (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention relates to a remote sensing image unsupervised change detection method based on a convolution self-codec, and belongs to the technical field of remote sensing images. The method comprises the following steps: s1, cutting, zooming and preprocessing remote sensing images of two different periods in the same place into image blocks, and dividing a data set; s2, respectively inputting the image blocks in two periods into the constructed remote sensing image change detection model, firstly coding to generate a multilayer feature map, then decoding to generate a multilayer feature map, and selecting two or more pairs of feature maps in two periods corresponding to the two or more layers of feature maps generated by decoding to be sent into a change map predictor for predicting change; s3, calculating characteristic value loss and distribution loss in the characteristic graphs in different periods, obtaining total loss through summing, and training an optimization model according to the loss values; and S4, sending the remote sensing images in different periods into the model by using the final model to obtain a change prediction image. The invention greatly reduces the manpower and material resources required by the ground object class marking and simultaneously improves the precision of the change detection algorithm.
Description
Technical Field
The invention relates to a remote sensing image change detection method, in particular to a remote sensing image unsupervised change detection method based on a convolution self-codec, and belongs to the technical field of convolution neural networks and remote sensing images.
Background
The remote sensing image has very important application value in military and national economy. A large number of high-resolution remote sensing images are widely used in tasks such as city planning, ground object classification, and the like. The change detection is a key and difficult point in the remote sensing field, and can be applied to various fields such as agriculture, civil use, military and the like. A plurality of units and scholars at home and abroad use the deep learning technology in a change detection project of the remote sensing image, and the detection precision and efficiency of extracting the ground object target change from the remote sensing image are effectively improved.
The change detection method based on deep learning comprises a supervision type and an unsupervised type, the two types of methods are both based on a deep convolutional neural network, the characteristics of remote sensing images in two periods are extracted through the deep convolutional neural network, change detection results are obtained by means of a supervision or unsupervised loss function, multiple research results show the advantages of the deep learning method compared with the traditional method, and the deep learning technology is more and more favored in the field of change detection of remote sensing images. The high-resolution image has the characteristics of multiple contents and large size, a large amount of manpower and material resources are needed for manufacturing the remote sensing image ground object classification data set, label missing and label error can be generated inevitably in the labeling process, and the prediction precision of the deep learning algorithm can be influenced. Compared with the supervised method, the unsupervised change detection method does not need accurate supervised information for auxiliary training, and meanwhile, due to the lack of the supervised information, the detection precision of the unsupervised method is difficult to avoid and has a defect, so that the advantage that the unsupervised change detection method does not need the supervised information is maintained, and the improvement of the precision of extracting the ground object target change detection is a very challenging problem.
At present, a convolutional self-encoder is applied to the field of remote sensing image change detection by scholars, remote sensing images in two periods are encoded, and ground feature characteristics in the remote sensing images are extracted to obtain a change map. The patent CN 114926512A discloses a twin convolution network remote sensing change detection method based on a fitting exclusive or function, which comprises the steps of obtaining a remote sensing image registered in two periods; inputting the two-stage remote sensing image into a trained twin convolution change detection network of the fitting exclusive or function, and outputting a change intensity graph through the twin convolution change detection network of the fitting exclusive or function; the twin convolution change detection network for fitting the exclusive or function comprises a twin encoder, an exclusive or module and a decoder, a remote sensing image is input into the twin encoder, and a multi-scale feature image is output through the twin encoder; inputting the multi-scale feature image into an XOR module, and outputting an initial change feature image through the XOR module; inputting the initial change characteristic image into a decoder, and outputting a change intensity graph through the decoder; the change detection result is obtained based on threshold extraction. The method uses remote sensing images and change truth labels in two periods, uses a cross entropy loss function to calculate loss, belongs to a typical supervised learning method, needs the change truth label manually labeled, and greatly increases manpower and material resources required by surface feature class labeling.
Disclosure of Invention
The invention aims to overcome the defects and provide the unsupervised change detection method of the remote sensing image based on the convolution self-codec, the remote sensing images in two periods are only needed, the change truth value label of manual annotation is not needed, the pixel similarity and the distribution difference training model of the characteristic diagram are obtained after the remote sensing images in the two periods pass through the convolution self-codec, the manpower and material resources required by ground feature class annotation are greatly reduced, and the precision of the change detection algorithm is improved.
The technical scheme adopted by the invention is as follows:
a remote sensing image unsupervised change detection method based on a convolution self-codec comprises the following steps:
s1, cutting, zooming and preprocessing remote sensing images of two different periods at the same place into image blocks, and dividing a data set;
s2, respectively inputting the image blocks in two periods into the constructed remote sensing image change detection model, firstly coding to generate a multilayer feature map, then decoding to generate a multilayer feature map, and selecting two or more pairs of feature maps in two periods corresponding to the two or more layers of feature maps generated by decoding and sending the two or more pairs of feature maps into a change map predictor to predict changes;
the remote sensing image change detection model comprises a convolution self-encoder and a change diagram predictor, wherein the convolution self-encoder is composed of an encoder and a decoder, the encoder is composed of a plurality of encoding modules, the decoder is composed of a plurality of decoding modules, the encoder generates feature diagrams with a plurality of sizes, namely multilayer feature diagrams, the decoder decodes the last layer of feature diagram and sequentially generates the feature diagrams with a plurality of sizes, namely multilayer feature diagrams;
s3, calculating the characteristic value loss in the characteristic diagrams in different periods by using the error square sum, calculating the distribution loss of the characteristic diagrams in different periods by using the Kullback-Leibler divergence, obtaining the total loss by summing, and training an optimization model according to the loss value;
and S4, sending the remote sensing images in different periods into the model by using the finally trained change detection model to obtain a change prediction image.
In the method, the pretreatment of the step S1 is to randomly select a part of the image blocks to carry out up-down, left-right and immediate overturning; calculating the mean value and standard deviation of RGB three channels, and normalizing the blocks to obtain blocks with the size of 256 × 3.
Each coding module in step S2 comprises a 3 × 3 convolution, a batch normalization layer, and a ReLU activation function layer. The batch normalization layer is used for normalizing numerical values in one batch, and helps to accelerate the training process of the model and reduce overfitting phenomena caused by regularization. The ReLU activation function layer changes all negative values to 0 to increase the non-linearity. The encoder preferably comprises four encoding modules, generates characteristic diagrams with four sizes, and two periods respectively correspond to the characteristic diagramsF 1 1 AndF 2 1 、F 1 2 andF 2 2 、F 1 3 andF 2 3 、F 1 4 andF 2 4 . Each decoding module contains a 3 x 3 deconvolution, a batch normalization layer, and a ReLU activation function layer. The decoder preferably comprises three decoding modules, decodes the feature map of the last layer generated by encoding, sequentially generates three-size feature maps, and respectively corresponds to the feature maps in two periodsF 1 1’ AndF 2 1’ 、 F 1 2’ andF 2 2’ 、 F 1 3’ andF 2 3’ 。
the change map predictor in the step S2 comprises a 1 × 1 convolution, a 3 × 3 convolution and a sigmoid function, wherein the 1 × 1 convolution is used for reducing the dimension of a channel, reducing the dimension of the feature map input into the discriminator to 256 dimensions, and reducing the memory consumption during calculation loss. The effect of the 3 x 3 convolution is to increase the non-linearity. The Sigmoid function normalizes the pixel values to [0,1]In the meantime. Preferably will result inF 1 2’ AndF 2 2’ 、 F 1 3’ andF 2 3’ two pairs of profiles are input to the profile predictor.
The calculation formula for calculating the eigenvalue loss in the feature map in different periods by using the sum of squared errors in step S3 is as follows:
wherein, N is the number of layers of the feature map participating in calculation, 2, X is the feature map of the remote sensing image in the period 1, X 'is the feature map of the remote sensing image in the period 2, and the difference value is calculated for each pair of pixels in the two images of X and X';
the calculation formula for calculating the distribution loss of the characteristic diagrams in different periods by using the Kullback-Leibler divergence is as follows:
wherein, the first and the second end of the pipe are connected with each other,
is the feature map probability distribution for epoch 1,
is the profile probability distribution for epoch 2.
The loss function of step S3 includes two parts: the feature similarity loss and the feature distribution loss calculate the difference between two period feature maps from two layers of feature values and feature distribution, so that the feature difference calculation of the two periods is more reasonable, and the coding and decoding capabilities of the convolutional self-codec are further improved.
And S3, model training optimization is carried out according to the loss value, an SGD optimizer is used for training, the momentum is 0.9, the learning rate is gradually increased to 0.01 by adopting a preheating method when training is started, the batch size is set to be 4, namely, one iteration is carried out on four pairs of pictures during training, each pair of pictures is a remote sensing image of two periods in the same place, four ten thousand iterations are carried out totally, and loss is output after each iteration is finished. And testing once after the training set is iterated once, and outputting the testing precision so as to predict the training degree of the neural network.
Another object of the present invention is to provide an unsupervised change detection device for remote sensing images based on convolutional self-codec, which comprises a memory, a processor and a computer program stored in the memory and running on the processor, wherein the processor executes the program to implement the unsupervised change detection method for remote sensing images based on convolutional self-codec.
The invention also provides a storage device which is a computer readable storage device, and the computer readable storage device is stored with a computer program for implementing the steps of the method for detecting unsupervised change of remote sensing images based on the convolution self-codec.
The invention has the beneficial effects that:
the method comprises the steps of encoding and decoding the remote sensing images in two periods by means of a convolution self-codec to generate feature maps in multiple sizes, and then calculating the similarity and distribution difference of the generated feature maps of the remote sensing images in different periods respectively, so as to train a change detection model. The invention relates to an unsupervised learning method based on error square sum loss and distribution loss, which only needs remote sensing images in two periods, does not need a change truth value label of manual annotation and does not need a ground object class annotation map of the remote sensing images, thereby greatly saving the annotation cost. Meanwhile, KL divergence is introduced to calculate distribution difference of the characteristics in different periods, distribution loss can be restrained on the characteristic distribution layer surface for characteristic graphs in two periods, the capability of the model for extracting the characteristics of the remote sensing images in the two periods is further improved, and therefore the accuracy of the change detection algorithm is further improved.
Drawings
FIG. 1 is a flow chart of the method of the present invention;
FIG. 2 is a schematic diagram of a model structure of the method of the present invention;
FIG. 3 is a block diagram of a convolutional self-codec of the present invention;
FIG. 4 is a block diagram of an encoding module of the present invention;
FIG. 5 is a block diagram of a decoding module according to the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is further described in detail with reference to the following embodiments and the accompanying drawings.
Example 1
A remote sensing image unsupervised change detection method based on a convolution self-codec comprises the following steps:
s1, cutting, zooming and preprocessing remote sensing images of two different periods at the same place into image blocks, and dividing a data set:
and cutting the remote sensing images in the two periods to form the remote sensing images in different periods in the same place. Scaling the cut remote sensing image to obtain 256 × 256 image blocks; randomly selecting part of the picture blocks to be turned over up, down, left and right; calculating the mean value and standard deviation of RGB three channels, and normalizing the blocks to obtain the final training block size of 256 × 3 (length × width × number of channels). All pictures were taken as 7: the ratio of 3 is randomly divided into a training set and a test set.
S2, respectively inputting the image blocks in two periods into the constructed remote sensing image change detection model, firstly coding to generate a multilayer characteristic diagram, then decoding to generate a multilayer characteristic diagram, and selecting two pairs of characteristic diagrams in two periods corresponding to the two layers of characteristic diagrams generated by decoding and sending the two pairs of characteristic diagrams into a change diagram predictor to predict changes:
referring to fig. 2, the remote sensing image change detection model comprises a convolutional self-codec and a change map predictor, the convolutional self-codec is composed of an encoder and a decoder, the encoder is composed of four encoding modules, and each encoding module (shown in fig. 4) comprises 3 × 3 convolution, a batch normalization layer and a ReLU activation function layer. The batch normalization layer is used for normalizing the values in one batch, and helps to accelerate the training process of the model and reduce overfitting phenomena caused by regularization. The ReLU activation function layer changes all negative values to 0 to increase the non-linearity. The decoder consists of three decoding modules, each (as shown in fig. 5) containing a 3 x 3 deconvolution, a batch normalization layer, and a ReLU activation function layer. The encoder generates feature maps of four sizes, and the decoder decodes the feature map of the last layer to sequentially generate feature maps of 3 sizes.
The blocks in the training set 256 × 3 are sent to a convolutional self-codec (fig. 3), and a four-layer feature map is obtained through an encoder: firstly, inputting image blocks of two periods into a first coding module of a decoder, and generating feature maps of the two periods through 3-by-3 convolution, batch normalization and ReLU activation in sequenceF 1 1 AndF 2 1 size 64 × 256; then the generated feature map is usedF 1 1 AndF 2 1 inputting the data into a second encoding module, performing the same operation as the above steps, sequentially performing 3-by-3 convolution, batch normalization and ReLU activation to generate feature maps of two periodsF 1 2 AndF 2 2 size 32 x 512; in the same way, the generated characteristic diagram is usedF 1 2 AndF 2 2 inputting the data into a third coding module, sequentially performing 3-by-3 convolution, batch normalization and ReLU activation to generate feature maps of two periodsF 1 3 AndF 2 3 size 16 x 1024; finally, the feature map is processedF 1 3 AndF 2 3 inputting the data into a fourth coding module, sequentially performing 3-by-3 convolution, batch normalization and ReLU activation to generate feature maps of two periodsF 1 4 AndF 2 4 and size 8 x 2048.
The convolutional self-encoder generates four layers of feature maps with the sizes respectively as follows: 64 × 256, 32 × 512, 16 × 1024, 8 × 2048. Then, the feature map with the size of 8 × 2048 is sent to a decoder, and the decoder decodes the feature map to obtain a three-layer feature map: firstly, the image blocks of two periods are input into a first decoding module of a decoder, and are subjected to 3-by-3 deconvolution, batch normalization and ReLU activation sequentially to generate feature maps of the two periodsF 1 1’ AndF 2 1’ size 16 x 1024; then the generated feature map is usedF 1 1’ AndF 2 1’ inputting the data into a second decoding module, performing the same operation as the above steps, sequentially performing 3-by-3 deconvolution, batch normalization and ReLU activation to generate a feature map of two periodsF 1 2’ AndF 2 2’ size 32 x 512; in the same way, the generated feature map is usedF 1 2’ AndF 2 2’ inputting the data into a third decoding module, sequentially performing 3-by-3 deconvolution, batch normalization and ReLU activation to generate feature maps of two periodsF 1 3’ AndF 2 3’ and size 64 x 256.
Finally, the convolutional self-codec encodes the images of the two epochs, and the decoding operation sequentially generates the feature maps with the sizes of 16 × 1024, 32 × 512, 64 × 256.
The change diagram predictor comprises a 1 × 1 convolution, a 3 × 3 convolution and a sigmoid function, wherein the 1 × 1 convolution is used for reducing the dimension of a channel, reducing the dimension of a feature diagram input into the discriminator to 256 dimensions, and reducing the video memory consumption during calculation loss. The effect of the 3 x 3 convolution is to increase the non-linearity. The Sigmoid function normalizes the pixel values to between 0, 1.
S3, calculating the loss of the characteristic value in the characteristic diagram in different periods by using the sum of squared errors, calculating the distribution loss of the characteristic diagram in different periods by using Kullback-Leibler divergence, obtaining the total loss by summing, and training an optimization model according to the loss value:
the remote sensing image characteristic diagrams in different periods are sent to a change diagram predictor, and the remote sensing image characteristic diagrams are selected and obtained by the inventionF 1 2’ AndF 2 2’ 、 F 1 3’ andF 2 3’ two pairs of feature maps are input (in deep learning, the feature map obtained at the beginning has high resolution but insufficient semantic features, after multilayer convolution, the semantic features are gradually sufficient, but the resolution is reduced, so the invention balances the semantic features and the resolution, uses the 2 < rd > and 3 < rd > layer feature maps for calculating loss), and calculates error square sum loss and distribution loss for the two pairs of feature maps respectively. Due to the fact thatThe invention is an unsupervised training method, so that in the training stage, the loss between two pairs of feature maps is directly calculated without obtaining a change map, and the model is optimized; in the testing stage, the value range of the feature map after sigmoid normalization is [0,1]]Therefore, the two pairs of feature maps are directly subtracted to obtain a value range of [0,1]]Setting the threshold value T to 0.6, the pixel value>0.6 is considered to be changed and is set to 1, and is set to 0 otherwise.
The invention uses the error square sum to calculate the loss of characteristic value in characteristic diagram in different periods, and uses Kullback-Leibler divergence to calculate the distribution loss of characteristic diagram in different periods, wherein the calculation formula of the error square sum is as follows:
wherein, N is the number of layers of the feature map participating in calculation, 2, X is the feature map of the remote sensing image in the period 1, X 'is the feature map of the remote sensing image in the period 2, and the difference value is calculated for each pair of pixels of each channel in the two images X and X';
the calculation formula for the Kullback-Leibler divergence is as follows:
wherein, the first and the second end of the pipe are connected with each other,
for the feature map probability distribution of epoch 1,
is the profile probability distribution for epoch 2.
In the change detection, the remote sensing images of the same place in two periods have differences due to various reasons, the invention uses the sum of squares of errors to calculate the difference of the pixel characteristic similarity of the two periods, and uses KL divergence to calculate the difference of the characteristic distribution of the two periods. The larger the change of the remote sensing images in two periods is, the larger the difference between the extracted characteristic maps is, and the larger the loss value calculated by means of the error sum of squares function and KL divergence is. Therefore, the sum of squared error calculated loss and the KL divergence calculated loss are added to obtain the total loss for back propagation.
And optimizing based on the loss function, training by using an SGD optimizer, wherein the momentum is 0.9, and gradually increasing the learning rate to 0.01 by adopting a preheating method when training is started. And setting the batch size to be 4, namely performing one iteration by four pairs of pictures during training, wherein each pair of pictures is a remote sensing image of the same place in two periods, and outputting loss after each iteration is finished. And testing once after the training set is iterated once, and outputting the testing precision so as to predict the training degree of the neural network.
S4, sending the remote sensing images in different periods into the model by using the finally trained change detection model to obtain a change prediction graph:
and after a finally trained model is obtained, sending the paired remote sensing images to be detected in different periods into the model, and obtaining a change prediction image.
Example 2
This embodiment provides a hardware device implementing the present invention:
a storage device is a computer readable storage device, and a computer program is stored on the computer readable storage device for implementing the steps in the method for unsupervised change detection of remote sensing images based on convolutional self-codec as described in embodiment 1.
An unsupervised change detection device for remote sensing images based on a convolutional self-codec comprises a memory, a processor and a computer program stored on the memory and capable of running on the processor, wherein the processor executes the program to realize the unsupervised change detection method for remote sensing images based on the convolutional self-codec in embodiment 1.
The above description is only exemplary of the present invention and should not be taken as limiting the invention, as any modification, equivalent replacement, or improvement made within the spirit and scope of the present invention should be included in the present invention.
Claims (7)
1. A remote sensing image unsupervised change detection method based on a convolution self-codec is characterized by comprising the following steps:
s1, cutting, zooming and preprocessing remote sensing images of two different periods at the same place into image blocks, and dividing a data set;
s2, respectively inputting the image blocks in two periods into the constructed remote sensing image change detection model, firstly coding to generate a multilayer characteristic diagram, then decoding to generate a multilayer characteristic diagram, and selecting two pairs of characteristic diagrams in two periods corresponding to the two layers of characteristic diagrams generated by decoding and sending the two pairs of characteristic diagrams into a change diagram predictor to predict changes;
the remote sensing image change detection model comprises a convolution self-encoder-decoder and a change graph predictor, wherein the convolution self-encoder-decoder consists of an encoder and a decoder, the encoder consists of four encoding modules and generates feature graphs with four sizes, and two periods respectively correspond to the feature graphsF 1 1 AndF 2 1 、F 1 2 andF 2 2 、F 1 3 andF 2 3 、F 1 4 andF 2 4 (ii) a The decoder is composed of three decoding modules, decodes the feature map of the last layer generated by encoding, sequentially generates three-size feature maps, and corresponds to the feature maps in two periodsF 1 1’ AndF 2 1’ 、 F 1 2’ andF 2 2’ 、 F 1 3’ andF 2 3’ (ii) a Will obtainF 1 2’ AndF 2 2’ 、 F 1 3’ andF 2 3’ inputting two pairs of feature maps into a change map predictor;
s3, calculating the characteristic value loss in the characteristic diagrams in different periods by using the error square sum, calculating the distribution loss of the characteristic diagrams in different periods by using the Kullback-Leibler divergence, obtaining the total loss by summing, and training an optimization model according to the loss value; the formula for calculating the loss of the eigenvalue in the characteristic diagram at different periods by using the sum of squared errors is as follows:
wherein N is the number of layers of the feature map involved in calculation, N is 2, X is the remote sensing image feature map of the period 1, X 'is the remote sensing image feature map of the period 2, and the difference value is calculated for each pair of pixels in the two maps X and X';
the calculation formula for calculating the distribution loss of the characteristic diagram in different periods by using the Kullback-Leibler divergence is as follows:
wherein the content of the first and second substances,
for the feature map probability distribution of epoch 1,
is the profile probability distribution for epoch 2;
and S4, sending the remote sensing images in different periods into the model by using the finally trained change detection model to obtain a change prediction graph.
2. The method for detecting the unsupervised change of the remote sensing image based on the convolutional self-codec as claimed in claim 1, wherein the preprocessing of the step S1 is to randomly select a part of image blocks to be turned over up, down, left and right; calculating the mean value and standard deviation of RGB three channels, and normalizing the blocks to obtain blocks with the size of 256 × 3.
3. The unsupervised change detection method of remote sensing images based on convolutional self-codec as claimed in claim 1, wherein each of the plurality of coding modules of step S2 comprises a 3 × 3 convolution, a batch normalization layer, a ReLU activation function layer; each decoding module of the plurality of decoding modules comprises a 3 × 3 deconvolution layer, a batch normalization layer and a ReLU activation function layer.
4. The unsupervised change detection method for the remote sensing image based on the convolutional self-codec as claimed in claim 1, wherein the change map predictor in step S2 comprises a 1 × 1 convolution, a 3 × 3 convolution and a sigmoid function.
5. The unsupervised change detection method of the remote sensing image based on the convolutional self-codec as claimed in claim 1, wherein in step S3, model training optimization is performed according to the loss value, an SGD optimizer is used for training, the momentum is 0.9, the learning rate is gradually increased to 0.01 by adopting a preheating method when training is started, the batch size is set to 4, that is, four pairs of pictures are calculated to perform one iteration during training, each pair of pictures is the remote sensing image of two periods in the same place, four million iterations are performed, and the loss is output after each iteration is completed.
6. An unsupervised change detection device for remote sensing images based on a convolutional self-codec, comprising a memory, a processor and a computer program stored on the memory and capable of running on the processor, characterized in that the processor implements the unsupervised change detection method for remote sensing images based on a convolutional self-codec according to any one of claims 1 to 5 when executing the program.
7. A storage device being a computer readable storage device having stored thereon a computer program for implementing the steps of the method for unsupervised change detection of remote sensing images based on convolutional self-codec as claimed in any one of claims 1 to 5.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211358970.XA CN115409842B (en) | 2022-11-02 | 2022-11-02 | Remote sensing image unsupervised change detection method based on convolution self-encoder-decoder |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211358970.XA CN115409842B (en) | 2022-11-02 | 2022-11-02 | Remote sensing image unsupervised change detection method based on convolution self-encoder-decoder |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN115409842A CN115409842A (en) | 2022-11-29 |
| CN115409842B true CN115409842B (en) | 2023-02-24 |
Family
ID=84169151
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211358970.XA Active CN115409842B (en) | 2022-11-02 | 2022-11-02 | Remote sensing image unsupervised change detection method based on convolution self-encoder-decoder |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115409842B (en) |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111291622A (en) * | 2020-01-16 | 2020-06-16 | 武汉汉达瑞科技有限公司 | Method and device for detecting building change in remote sensing image |
| CN112668494A (en) * | 2020-12-31 | 2021-04-16 | 西安电子科技大学 | Small sample change detection method based on multi-scale feature extraction |
| WO2021211787A1 (en) * | 2020-04-15 | 2021-10-21 | Children's Hospital Medical Center | Systems and methods for quantification of liver fibrosis with mri and deep learning |
| CN113936217A (en) * | 2021-10-25 | 2022-01-14 | 华中师范大学 | Priori semantic knowledge guided high-resolution remote sensing image weakly supervised building change detection method |
| CN114022793A (en) * | 2021-10-28 | 2022-02-08 | 天津大学 | Optical remote sensing image change detection method based on twin network |
| CN114283120A (en) * | 2021-12-01 | 2022-04-05 | 武汉大学 | End-to-end multi-source heterogeneous remote sensing image change detection method based on domain self-adaptation |
| CN114926512A (en) * | 2022-05-31 | 2022-08-19 | 武汉大学 | Twin convolution network remote sensing change detection method based on fitting exclusive or function |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP3952306A1 (en) * | 2018-10-02 | 2022-02-09 | Nokia Technologies Oy | An apparatus, a method and a computer program for running a neural network |
| CN114943672A (en) * | 2021-02-09 | 2022-08-26 | 富泰华工业（深圳）有限公司 | Image defect detection method and device, electronic equipment and storage medium |
-
2022
- 2022-11-02 CN CN202211358970.XA patent/CN115409842B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111291622A (en) * | 2020-01-16 | 2020-06-16 | 武汉汉达瑞科技有限公司 | Method and device for detecting building change in remote sensing image |
| WO2021211787A1 (en) * | 2020-04-15 | 2021-10-21 | Children's Hospital Medical Center | Systems and methods for quantification of liver fibrosis with mri and deep learning |
| CN112668494A (en) * | 2020-12-31 | 2021-04-16 | 西安电子科技大学 | Small sample change detection method based on multi-scale feature extraction |
| CN113936217A (en) * | 2021-10-25 | 2022-01-14 | 华中师范大学 | Priori semantic knowledge guided high-resolution remote sensing image weakly supervised building change detection method |
| CN114022793A (en) * | 2021-10-28 | 2022-02-08 | 天津大学 | Optical remote sensing image change detection method based on twin network |
| CN114283120A (en) * | 2021-12-01 | 2022-04-05 | 武汉大学 | End-to-end multi-source heterogeneous remote sensing image change detection method based on domain self-adaptation |
| CN114926512A (en) * | 2022-05-31 | 2022-08-19 | 武汉大学 | Twin convolution network remote sensing change detection method based on fitting exclusive or function |
Also Published As
| Publication number | Publication date |
|---|---|
| CN115409842A (en) | 2022-11-29 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109241972B (en) | Image semantic segmentation method based on deep learning | |
| CN115049936A (en) | High-resolution remote sensing image-oriented boundary enhancement type semantic segmentation method | |
| CN113592786B (en) | Deep learning-based ocean mesoscale vortex detection method | |
| CN111832484A (en) | Loop detection method based on convolution perception hash algorithm | |
| CN110930378B (en) | Emphysema image processing method and system based on low data demand | |
| CN112598053A (en) | Active significance target detection method based on semi-supervised learning | |
| CN115410059B (en) | Remote sensing image part supervision change detection method and device based on contrast loss | |
| CN113066089B (en) | Real-time image semantic segmentation method based on attention guide mechanism | |
| CN115984850A (en) | Lightweight remote sensing image semantic segmentation method based on improved Deeplabv3+ | |
| CN116383422B (en) | Non-supervision cross-modal hash retrieval method based on anchor points | |
| CN114241218A (en) | Target significance detection method based on step-by-step attention mechanism | |
| CN113870286A (en) | Foreground segmentation method based on multi-level feature and mask fusion | |
| CN115565056A (en) | Underwater image enhancement method and system based on condition generation countermeasure network | |
| CN115147819A (en) | Driver fixation point prediction method based on fixation point prediction model | |
| CN114154016A (en) | Video description method based on target space semantic alignment | |
| CN114299305B (en) | Saliency target detection algorithm for aggregating dense and attention multi-scale features | |
| CN115331073A (en) | Image self-supervision learning method based on TransUnnet architecture | |
| CN114882434A (en) | Unsupervised abnormal behavior detection method based on background suppression | |
| CN115908793A (en) | Coding and decoding structure semantic segmentation model based on position attention mechanism | |
| CN114612803A (en) | Transmission line insulator defect detection method for improving CenterNet | |
| Tian et al. | Semantic segmentation of remote sensing image based on GAN and FCN network model | |
| CN117765258A (en) | Large-scale point cloud semantic segmentation method based on density self-adaption and attention mechanism | |
| CN115409842B (en) | Remote sensing image unsupervised change detection method based on convolution self-encoder-decoder | |
| CN117152438A (en) | Lightweight street view image semantic segmentation method based on improved deep LabV3+ network | |
| CN117152823A (en) | Multi-task age estimation method based on dynamic cavity convolution pyramid attention |
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 |