Training method for detecting image change model and image change detection method

Abstract

The present invention discloses a training method for a model for detecting image changes, including: obtaining a training sample image and preprocessing the training sample image; using a generator network to process the first training sample image to obtain a first category probability map; performing a heat map on the second training sample image to obtain a second category probability map; using the first training sample image and the second training sample image to obtain a difference map; respectively operating the difference map with the first category probability map and the second category probability map to obtain a first sample image and a second sample image; using a discriminant perception network to process the first sample image and the second sample image to obtain a discrimination result; using the discrimination result and the first category probability map, according to a loss function, optimizing the parameters of the generator network and the parameters of the discriminant perception network to obtain a trained generator network and a trained discriminant perception network. The present invention also discloses an image change detection method.

Description

Technical Field

[0001] The present invention relates to the field of image detection, in particular to the field of remote sensing image interpretation, and specifically to a training method for a model for detecting image changes and an image change detection method. Background Art

[0002] With the further development of sensor technology, aviation and space platform data communication, and the effective combination with ultra-low altitude unmanned aerial vehicles and vehicles and other platform technologies, and accompanied by the advancement of computer technology, remote sensing technology has achieved rapid development. As the main form of remote sensing data recording, remote sensing images are developing towards the "three highs" trend of high spatial resolution, high spectral resolution and high temporal resolution, providing time series observation data for all periods of time in all regions of the world. The accumulation of massive multi-temporal remote sensing data provides an important data source for global change research such as land cover change detection at regional and even global scales. Remote sensing image change detection refers to the use of multi-source remote sensing images and related geographic data covering the same surface area in different historical periods, combined with the corresponding land feature characteristics and remote sensing imaging mechanism, using image graphics processing theory and mathematical model methods to determine and analyze the changes in the land features in the region, including changes in the location range of the land features and changes in the nature and state of the land features.

[0003] In recent years, the emergence of deep learning has brought a series of revolutionary advances in the field of image interpretation, including change detection technology. Relying on the powerful feature extraction capabilities of convolutional neural networks, change detection has changed from manual feature extraction to an automatic integrated process. However, the convolutional neural network methods currently used generally only use labels for optimization, without fully considering the high-order consistency between the output of the change detection model and the true value label, and the perception ability of the changed area is weak. Summary of the invention

[0004] In order to solve the above problems, the present invention provides a training method for a model for detecting image changes and an image change detection method.

[0005] According to one aspect of the present invention, a method for training a model for detecting image changes is provided, comprising:

[0006] Acquire a training sample image and preprocess the training sample image, wherein the training sample image includes a first training sample image and a second training sample image;

[0007] Processing the first training sample image using the generator network to obtain a first category probability map of the first training sample image;

[0008] Performing heat mapping on the second training sample image to obtain a second category probability map of the second training sample image;

[0009] Obtaining a difference map using the first training sample image and the second training sample image;

[0010] The difference map is respectively operated with the first category probability map and the second category probability map to obtain a first sample image and a second sample image;

[0011] Processing the first sample image and the second sample image using a discriminative perception network to obtain a discriminative result;

[0012] Using the discrimination result and the first category probability map, according to the loss function, the parameters of the generator network and the parameters of the discriminant perception network are optimized to obtain a trained generator network and a trained discriminant perception network.

[0013] According to an embodiment of the present invention, the preprocessing of the training sample image includes:

[0014] The training sample image is cut multiple times using a sliding window algorithm to obtain multiple cut images;

[0015] The cut images are inverted at multiple angles to obtain enhanced training sample images.

[0016] According to an embodiment of the present invention, the above-mentioned processing of the first training sample image by using the generator network to obtain the first category probability map of the first training sample image includes:

[0017] Construct a generator network and initialize the parameters of the generator network. The generator network includes an encoder layer and a decoder layer.

[0018] Performing a cascade operation on the first training sample image, and inputting the cascaded first training sample image into the generator network;

[0019] Down-sampling the cascaded first training sample image through an encoder layer to obtain down-sampled feature data;

[0020] Upsampling the cascaded first training sample image through a decoder to obtain upsampled feature data;

[0021] The upsampled data and feature data are fused to obtain the first category probability map.

[0022] According to an embodiment of the present invention, the heat mapping of the second training sample image to obtain the second category probability map of the second training sample image includes:

[0023] According to a preset range, obtain the true value labels of other pixel points centered on the pixel point to be processed, where the preset range is a rectangular area;

[0024] According to the number of pixels in the rectangular area and the true value labels of the pixels, the thermal value of the processed pixels is calculated;

[0025] The pixels of the second training sample image are traversed to obtain a second category probability map.

[0026] According to an embodiment of the present invention, the above-mentioned use of the discriminative perception network to process the first sample image and the second sample image to obtain a discrimination result includes:

[0027] Constructing a discriminative perception network and initializing parameters of the discriminative perception network, wherein the discriminative perception network includes a convolutional layer and a fully connected layer;

[0028] Randomly inputting the first sample image and the second sample image into the discriminative perception network;

[0029] When the first sample image is input, the discrimination result is 0;

[0030] When the second sample image is input, the discrimination result is 1.

[0031] According to an embodiment of the present invention, the above-mentioned use of the discrimination result and the first category probability map to optimize the parameters of the generator network and the parameters of the discriminant perception network according to the loss function to obtain the trained generator network and the trained discriminant perception network includes:

[0032] The generator network optimizes the parameters of the generator network according to the weighted contrast loss function using the discrimination result;

[0033] Use the optimized generator network to process the training sample images and obtain the category probability map;

[0034] The discriminant perception network optimizes the discriminant perception network according to the cross entropy loss function using the category probability map;

[0035] The optimization operation is performed iteratively until the values ​​of the weighted contrast loss function and the cross entropy loss function meet the preset conditions, and a trained generator network and a trained discriminant perception network are obtained.

[0036] According to an embodiment of the present invention, the weighted contrast loss function is expressed by formula (1):

[0037]

[0038] Where W represents the width of the first training sample image, H represents the height of the first training sample image, and w c and w u is the weight of the changed sample and the unchanged sample, y ij is the true value label of the first training sample image, p ij is the output probability value of the generator network.

[0039] According to an embodiment of the present invention, the cross entropy loss function is expressed by formula (2):

[0040] l bgc (p,y)=-[ylnp+(1-y)ln(1-p)] (2),

[0041] Among them, p represents the probability value output by the discriminator, and y is the true value label of the discriminator.

[0042] According to an embodiment of the present invention, the above loss function includes a weighted contrast loss function and a cross entropy loss function, and the loss function is expressed by formula (3):

[0043]

[0044] Among them, WCL pij is the weighted contrast loss function, γ is the weight parameter, x ij is the pixel corresponding to the cascade input of the two-phase image, l bec is the cross entropy loss function, diff(x ij ) is the pixel corresponding to the difference map of the two phase images, h represents the thermal mapping process, y ij It is the corresponding label in the heat map process.

[0045] According to another aspect of the present invention, there is provided an image change detection method, comprising:

[0046] Acquire an image to be detected;

[0047] Constructing a generator network, and using the generator network to process the image to be detected, wherein the generator network is trained by the above training method;

[0048] Obtaining a discrimination result of the image to be detected.

[0049] The training method of the model for detecting image changes provided by the present invention can fully consider the high-order consistency problem between the output of the change detection model and the true value label, and the similarity between the detection result obtained by the perception model and the real change result; at the same time, the training method of the above model can effectively extract the features of the image by adding a discriminant perception network, thereby improving the performance of the trained model. In addition, the image change detection method provided by the present invention makes full use of the detection model obtained by the above training method, and can be applied in the field of optical remote sensing image change detection to improve the accuracy and robustness of change detection of optical remote sensing images. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 A flow chart of a training method for a model for detecting image changes according to an embodiment of the present invention is schematically shown;

[0051] Figure 2 A flowchart of preprocessing training sample images according to an embodiment of the present invention is schematically shown;

[0052] Figure 3 Schematically shows a flow chart of obtaining a first category probability map according to an embodiment of the present invention;

[0053] Figure 4 Schematically shows a flow chart of obtaining a second category probability map according to an embodiment of the present invention;

[0054] Figure 5 is a schematic diagram of a heat map of true value labels according to an embodiment of the present invention;

[0055] Figure 6 is a flowchart of obtaining a trained generator network and a trained discriminant perception network according to an embodiment of the present invention;

[0056] Figure 7 is a schematic structural diagram of a model for detecting image changes according to an embodiment of the present invention;

[0057] Figure 8 is a flow chart of an image change detection method according to an embodiment of the present invention;

[0058] Fig. 9 is an example diagram of an image to be detected according to an embodiment of the present invention;

[0059] Fig.10 It is a schematic diagram of the result of processing an image with detection using the image change detection method of the present invention. DETAILED DESCRIPTION

[0060] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with specific embodiments and with reference to the accompanying drawings.

[0061] Figure 1 The flowchart of the training method of the model for detecting image changes according to an embodiment of the present invention is schematically shown.

[0062] like Figure 1 As shown, operation S110 to operation S170 are included.

[0063] In operation S110, a training sample image is acquired and preprocessed, wherein the training sample image includes a first training sample image and a second training sample image;

[0064] The training sample images used in the embodiment of the present invention are SZTAKI aerial photography public data sets, which include 12 pairs of high-resolution optical aerial photography images with a resolution of 1.5 m. The first training sample image and the second training sample image are two-phase images.

[0065] In operation S120, a first training sample image is processed using a generator network to obtain a first category probability map of the first training sample image;

[0066] The generator network adopts the UNet network architecture. By modifying the parameters of the UNet network, the generator network provided by the present invention can receive two cascaded images.

[0067] In operation S130, a second training sample image is converted into a heat map to obtain a second category probability map of the second training sample image;

[0068] By performing heat mapping on the true value labels of the second training sample image, the form of the second category probability map is consistent with the first category probability map output by the generator network.

[0069] In operation S140, a difference map is obtained using the first training sample image and the second training sample image;

[0070] The first training sample image is subtracted from the second training sample image to obtain a difference map between the two.

[0071] In operation S150, the difference map is respectively operated with the first category probability map and the second category probability map to obtain a first sample image and a second sample image;

[0072] The first category probability image is multiplied by the difference map to obtain a first sample image. Similarly, the second category probability image is multiplied by the difference map to obtain a second sample image.

[0073] In operation S160, the first sample image and the second sample image are processed by using a discriminative perception network to obtain a discriminative result;

[0074] In operation S170, the parameters of the generator network and the parameters of the discriminant perception network are optimized according to the loss function using the discrimination result and the first category probability map, so as to obtain a trained generator network and a trained discriminant perception network.

[0075] The output of the discriminant perception network is transmitted back to optimize the generator network, and the output of the generator network is used to train the discriminant perception network so that the output of the discriminant perception network can better predict the output of the generator network. When the loss function value converges or, the discriminant perception network and the generator network reach the optimal state.

[0076] The above-mentioned training method for the model for detecting image changes fully processes the image to be detected, and uses the loss function to optimize the generator network and the discriminant perception network multiple times, so as to obtain a trained image detection model; at the same time, the application of this model can effectively detect changes in optical remote sensing images, and the model can also be used for image change detection in other fields.

[0077] Figure 2 The flowchart of training sample image preprocessing according to an embodiment of the present invention is schematically shown.

[0078] like Figure 2 As shown, operation S210 to operation S220 are included.

[0079] In operation S210, the training sample image is cut multiple times using a sliding window algorithm to obtain multiple cut images;

[0080] In operation S220, the cut image is inverted at multiple angles to obtain an enhanced training sample image.

[0081] The training sample images of the embodiment of the present invention adopt the SZTAKI aerial photography public data set, and the upper left part 784×448 of the image in the public data set is used as the test set, and the rest is used as the training set. The training set part uses a sliding window to repeatedly cut, and the cut images are inverted by 90 degrees, 180 degrees, 270 degrees and horizontally to enhance the data set.

[0082] By processing the training sample images through the above preprocessing process, the efficiency of the model training process can be effectively improved.

[0083] Figure 3 The flowchart for obtaining the first category probability map according to an embodiment of the present invention is schematically shown.

[0084] like Figure 3 As shown, including operation S310 to operation S350

[0085] In operation S310, a generator network is constructed and parameters of the generator network are initialized, where the generator network includes an encoder layer and a decoder layer;

[0086] In operation S320, a cascade operation is performed on the first training sample image, and the cascaded first training sample image is input into the generator network;

[0087] The cascade operation is aimed at processing two two-phase images. For example, in a two-phase image, each phase is an RGB three-channel image. The cascade is to cascade the images in the channel dimension, and the dimension sent to the network is six channels. If the input image of phase one is (3, w, h), 3 represents the RGB three channels, and wh represents the image width and height respectively, and the image of phase two is (3, w, h), the result after cascading is (6, w, h).

[0088] In operation S330, down-sampling the cascaded first training sample image through an encoder layer to obtain down-sampled feature data;

[0089] In operation S340, up-sampling the cascaded first training sample image is performed through a decoder to obtain up-sampled feature data;

[0090] In operation S350, the upsampled data and the feature data are fused to obtain a first category probability map.

[0091] The generator network performs convolution operations on the first training sample image input for each layer. The length and width of the features are halved and the number of channels is doubled. The feature channels output after the original image is convolved for the first time are 16. The network contains four encoding layers. The subsequent decoding structure is cascaded with the output of the previous encoding layer as the decoding output of the next layer. The final output of the network is activated by sigmoid to become a category probability map. The Sigmoid function is as follows:

[0092] Figure 4 The flowchart of obtaining the second category probability map according to an embodiment of the present invention is schematically shown.

[0093] like Figure 4 As shown, operation S410 to operation S430 are included.

[0094] In operation S410, the true value labels of other pixel points centered on the pixel point to be processed are obtained according to a preset range, wherein the preset range is a rectangular area;

[0095] In operation S420, the thermal value of the processed pixel is calculated according to the number of pixels in the rectangular area and the true value labels of the pixels;

[0096] In operation S430, pixel points of the second training sample image are traversed to obtain a second category probability map.

[0097] Figure 5 is a schematic diagram of the heat map of the true value label according to an embodiment of the present invention. Figure 5 The above process of obtaining the second category probability map is further explained.

[0098] like Figure 5 As shown in , the true value label is a binary image of 0 or 1. The heat map process makes the pixel point in the center of the change area approach 1 or equal to 1, while the change probability value of the edge area is smaller. Figure 5 As shown, a rectangular box is taken around a pixel point (i, j) in the true value label, which can be 3×3, 5×5, 7×7, etc. In the present invention, 3×3 is taken, and then the number of change points n in this box is calculated, and the value of the pixel point (i, j) is Where N is the number of pixels in the obtained rectangular frame, for example, in the present invention, the value can be 9. In this way, the entire label is traversed to obtain a heat map of the true value, that is, the second category probability map mentioned above in the present invention.

[0099] According to an embodiment of the present invention, the above-mentioned use of the discriminative perception network to process the first sample image and the second sample image to obtain a discrimination result includes:

[0100] Constructing a discriminative perception network and initializing parameters of the discriminative perception network, wherein the discriminative perception network includes a convolutional layer and a fully connected layer;

[0101] Randomly inputting the first sample image and the second sample image into the discriminative perception network;

[0102] When the first sample image is input, the discrimination result is 0;

[0103] When the second sample image is input, the discrimination result is 1.

[0104] Figure 6 It is a flowchart of obtaining a trained generator network and a trained discriminant perception network according to an embodiment of the present invention.

[0105] like Figure 6 As shown, operations S610 to S640 are included.

[0106] In operation S610, the generator network optimizes parameters of the generator network using the discrimination result according to a weighted contrast loss function;

[0107] In operation S620, the training sample image is processed using the optimized generator network to obtain a class probability map;

[0108] In operation S630, the discriminative perception network optimizes the discriminative perception network using the class probability map according to a cross entropy loss function;

[0109] In operation S640, the optimization operation is iteratively performed until the values ​​of the weighted contrast loss function and the cross entropy loss function meet preset conditions, thereby obtaining a trained generator network and a trained discriminant perception network.

[0110] The above method aims to obtain a trained generator network and a discriminant perception network, by randomly selecting the output of the generator or the heat map of the true value label and multiplying it with the difference map of the two input images as the input of the discriminator. The discriminator is a binary classification network. If the input is a combination of the true value label and the difference map, the discriminator label is 1, otherwise it is 0. The discriminator is a three-layer convolution structure and a fully connected layer. The number of input channels is 3. The number of output channels of the first convolution layer is 16, and the feature size is halved; the next two convolution layers are similar; the output of the fully connected layer is a binary classification, and the loss function uses cross entropy. If the input is a cascade of the true value label and the difference map, the discriminator label is 1, otherwise it is 0. When optimizing the network, the generator and discriminator are repeatedly updated until the network converges.

[0111] According to an embodiment of the present invention, the weighted contrast loss function is expressed by formula (1):

[0112]

[0113] Where W represents the width of the first training sample image, H represents the height of the first training sample image, and w c and w u is the weight of the changed sample and the unchanged sample, y ij is the true value label of the first training sample image, p ij is the output probability value of the generator network.

[0114] According to an embodiment of the present invention, the cross entropy loss function is expressed by formula (2):

[0115] l bec (p,y)=-[y lnp+(1-y)ln(1-p)] (2),

[0116] Among them, p represents the probability value output by the discriminator, and y is the true value label of the discriminator.

[0117] According to an embodiment of the present invention, the above loss function includes a weighted contrast loss function and a cross entropy loss function, and the loss function is expressed by formula (3):

[0118]

[0119] Among them, WCL pij is the weighted contrast loss function, γ is the weight parameter, x ij is the pixel corresponding to the cascade input of the two-phase image, l bec is the cross entropy loss function, diff(x ij ) is the pixel corresponding to the difference map of the two phase images, h represents the thermal mapping process, y ij It is the corresponding label in the heat map process.

[0120] The above loss function can improve the speed and accuracy of the training process. During the training process of the above model, when the generator network is optimized, the discriminator parameters are fixed, and the generator is optimized using the above loss function to obtain the optimized generator network; then the discriminator is optimized using the generator network with optimized parameters according to the above loss function, and the update is repeated until the network converges.

[0121] Figure 7 is a schematic diagram of the structure of a model for detecting image changes according to an embodiment of the present invention. Figure 7 The above training method is further explained.

[0122] like Figure 7 As shown, the model for detecting image changes provided by the embodiment of the present invention mainly includes a generator network (i.e., the generator in the figure) and a discriminant perception network (i.e., the discriminator in the figure), wherein the generator includes an encoder and a decoder; the two phase images are cascaded and input into the generator, the encoder and decoder in the generator process the cascade image of the two phase images, and output a category probability map, and it is particularly noted that the above two phase images have true value labels; at the same time, the image with the true value label is converted into a heat map for subsequent processing; the two phase difference map is multiplied with the category probability map or the heat map, and input into the discriminator, the discriminator is processed, and the discrimination result is output, such as true or false (true is represented by 1, and false is represented by 0). In addition, in the model provided by the present invention, the discriminator is only used as an auxiliary tool in network training optimization. During network testing, the discriminator is not used, and only the generator is used for reasoning. In the input probability map, if the probability value is greater than 0.5, it is judged as a changed pixel point, otherwise it is judged as a non-changed pixel point, and the final change detection result can be obtained. That is, the discriminator can optionally add the existing mainstream change detection method to perceive the similarity between the detection results obtained by the model and the actual change results, so that the detection results and the true value labels have better high-order consistency, thereby improving network performance.

[0123] Figure 8 is a flowchart of an image change detection method according to an embodiment of the present invention.

[0124] like Figure 8 As shown, operation S810 to operation S830 are included.

[0125] Operation S810, acquiring an image to be detected;

[0126] Operation S820, constructing a generator network, and using the generator network to process the image to be detected, wherein the generator network is trained by the above training method;

[0127] Operation S830: Obtain a discrimination result of the image to be detected.

[0128] Fig. 9 9 is an example diagram of an image to be detected according to an embodiment of the present invention, wherein 9(a) is TISZADOB / 3 phase 1, and 9(b) is TISZADOB / 3 phase 2.

[0129] Fig.10 10(a) is a schematic diagram of the result of processing an image with detection using the image change detection method of the present invention, wherein 10(b) is a schematic diagram of the result of processing an image with detection using the image change detection method provided by the present invention, 10(c) is a schematic diagram of the result of processing an image with detection using the prior art method 1, and 10(d) is a schematic diagram of the result of processing an image with detection using the prior art method 2.

[0130] Combine the following Fig. 9 and Fig.10 The image change detection method provided by the present invention is further described.

[0131] like Fig. 9 As shown, Fig. 9 (a) and Fig. 9 (b) is a two-phase image, which represents the optical remote sensing images of the same area at different times. The above two phase images are processed using the image change detection method provided by the present invention to obtain Fig.10 (b) The result graph, reference standard graph (i.e. Fig.10 (a)), using the image change detection method provided by the present invention, a result that is highly consistent with the standard image can be obtained, and compared with the results obtained by other methods in the prior art, the image accuracy is obviously greatly improved. Therefore, the image change detection method provided by the embodiment of the present invention can take into account the high-order consistency between the output of the change detection model and the true value label, has high robustness and accuracy, and can be widely used in the field of interpretation of optical remote sensing images. At the same time, the image change detection method provided by the present invention can also be applied to other image interpretation fields, rather than being limited to the field of interpretation of remote sensing images.

[0132] The specific embodiments described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for training a model for detecting image changes, comprising: Acquire a training sample image and preprocess the training sample image, wherein the training sample image includes a first training sample image and a second training sample image; Processing the first training sample image using a generator network to obtain a first category probability map of the first training sample image; Performing heat mapping on the second training sample image to obtain a second category probability map of the second training sample image; Obtaining a difference map using the first training sample image and the second training sample image; Performing operations on the difference map, the first category probability map and the second category probability map respectively to obtain a first sample image and a second sample image; Processing the first sample image and the second sample image using a discriminative perception network to obtain a discriminative result; Utilizing the discrimination result and the first category probability map, according to the loss function, the parameters of the generator network and the parameters of the discriminant perception network are optimized to obtain a trained generator network and a trained discriminant perception network.

2. According to the training method of claim 1, preprocessing the training sample image comprises: Using a sliding window algorithm to perform multiple cutting on the training sample image to obtain multiple cutting images; The cut image is inverted at multiple angles to obtain an enhanced training sample image.

3. The method according to claim 1, wherein: The using a generator network to process the first training sample image to obtain a first category probability map of the first training sample image includes: Constructing the generator network and initializing the parameters of the generator network, wherein the generator network includes an encoder layer and a decoder layer; Performing a cascade operation on the first training sample image, and inputting the cascaded first training sample image into the generator network; Down-sampling the cascaded first training sample image through the encoder layer to obtain down-sampled feature data; Up-sampling the cascaded first training sample image by the decoder to obtain up-sampled feature data; The up-sampled data and the feature data are fused to obtain the first category probability map.

4. The method according to claim 1, wherein: Converting the second training sample image into a heat map to obtain a second category probability map of the second training sample image includes: According to a preset range, obtaining the true value labels of other pixel points centered on the pixel point to be processed, wherein the preset range is a rectangular area; Calculate the thermal value of the pixel to be processed according to the number of pixels in the rectangular area and the true value labels of the pixels; The pixels of the second training sample image are traversed to obtain the second category probability map.

5. The method according to claim 1, wherein: The using a discriminant perception network to process the first sample image and the second sample image to obtain a discriminant result includes: Constructing the discriminative perception network and initializing the parameters of the discriminative perception network, wherein the discriminative perception network includes a convolutional layer and a fully connected layer; Randomly inputting the first sample image and the second sample image into the discriminative perception network; When the first sample image is input, the discrimination result is 0; When the second sample image is input, the discrimination result is 1.

6. The method according to claim 1, wherein: Utilizing the discrimination result and the first category probability map, according to the loss function, optimizing the parameters of the generator network and the parameters of the discriminant perception network, and obtaining the trained generator network and the trained discriminant perception network comprises: The generator network optimizes the parameters of the generator network according to the weighted contrast loss function using the discrimination result; Processing the training sample image using the optimized generator network to obtain a category probability map; The discriminant perception network optimizes the discriminant perception network according to the cross entropy loss function using the category probability map; The optimization operation is iteratively performed until the values ​​of the weighted contrast loss function and the cross entropy loss function meet preset conditions, thereby obtaining a trained generator network and a trained discriminant perception network.

7. The method according to claim 6, wherein: The weighted contrast loss function is expressed by formula (1): Wherein, W represents the width of the first training sample image, H represents the height of the first training sample image, and w c and w u is the weight of the changed sample and the unchanged sample, y ij is the true value label of the first training sample image, p ij is the output probability value of the generator network.

8. The method according to claim 6, wherein: The cross entropy loss function is expressed by formula (2): l bec (p,y)=-[ylnp+(1-y)ln(1-p)] (2), Among them, p represents the probability value output by the discriminator, and y is the true value label of the discriminator.

9. The method according to claim 1, wherein: The loss function includes a weighted contrast loss function and a cross entropy loss function, and the loss function is expressed by formula (3): in, is the weighted contrast loss function, γ is the weight parameter, x ij is the pixel corresponding to the cascade input of the two-phase image, l bec is the cross entropy loss function, diff(x ij ) is the pixel corresponding to the difference map of the two phase images, h represents the thermal mapping process, y ij is the corresponding label in the heat mapping process.

10. A method for detecting image changes, comprising: Acquire an image to be detected; Constructing a generator network, and using the generator network to process the image to be detected, wherein the generator network is trained by any one of the methods described in claims 1-9; Obtaining a discrimination result of the image to be detected.

Images