Heterogenous remote sensing image change detection method and system

Abstract

The invention relates to the technical field of heterogenous remote sensing image change detection, in particular to a heterogenous remote sensing image change detection method and system. Performing stationary wavelet decomposition on the target heterogenous remote sensing image, and processing the decomposed target heterogenous remote sensing image by adopting an image translation network to obtain a translated image; respectively acquiring a wavelet domain difference image and a spatial domain difference image between the translated image and the target heterogeneous remote sensing image; according to the wavelet domain difference image and the spatial domain difference image, synthesizing a fusion difference image between the translation image and the target heterogenous remote sensing image; pixel points in the fused difference image are divided through threshold segmentation, and a change detection result of the target heterogeneous remote sensing image is obtained; according to the method, the high-quality difference image can be generated, so that the change area in the image can be accurately identified, and the accuracy of change detection is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of heterogeneous remote sensing image change detection, and particularly relates to a method and system for heterogeneous remote sensing image change detection. Background Art

[0002] With the rapid development of remote sensing technology, using data from different observation platforms to monitor ground dynamic changes has become an important way to study the evolution of the Earth's environment. Change detection is a process of identifying changes by comparing a set of remote sensing images obtained in the same geographical area but at different times, and has been widely applied in many fields, such as deforestation, urban planning, and disaster assessment. Traditional change detection methods mainly target homogeneous images, that is, the obtained images are from the same sensor, such as high-resolution (VHR) optical images, synthetic aperture radar (SAR) images, and hyperspectral images. However, due to the limitations of image imaging conditions and satellite revisit cycles, it is impossible to obtain homogeneous images in a timely manner. Therefore, using remote sensing images from different sensors, that is, heterogeneous images, for change detection has begun to receive great attention.

[0003] Since heterogeneous images usually have different statistical characteristics, it is difficult for them to be directly compared like homogeneous images to generate a difference image. To address the above problems, an intuitive solution is to convert heterogeneous images into homogeneous images with similar statistical characteristics, making them comparable and highlighting changes. In recent years, deep learning technology has developed rapidly and has been successfully applied to the field of remote sensing image change detection. Currently, most deep learning-based methods are based on adversarial learning or style transfer methods to eliminate the modal differences between heterogeneous images. For example: in the prior art, there are methods based on the difference of affinity matrices, namely X-Net and adversarial cycle encoder network (ACENet), which aim to combine cycle consistency and adversarial training to translate images. There is also an unsupervised change detection method based on autoencoders (CAAE) to force the alignment of the feature spaces of heterogeneous images and reduce the impact of pixel changes on image translation. There is also a method based on dual-image translation and dual-contrast learning in the prior art for accurately identifying surface changes. An end-to-end multi-domain constrained translation network is designed, which uses an enhanced generative adversarial network (GAN) to ensure the realism of the translated images and integrates contrast learning to retain the content information of the source domain images. Although the above methods have achieved good change detection results currently, they either need to make changes before guiding the translation process or require multiple rounds of training of GAN and parameter optimization of style transfer learning, resulting in the inability to capture subtle changes in complex scenes. Summary of the Invention

[0004] To solve the problems existing in the prior art, the present invention provides a method for detecting changes in heterogeneous remote sensing images. The method includes obtaining target heterogeneous remote sensing images; performing stationary wavelet decomposition on the target heterogeneous remote sensing images, and processing the decomposed target heterogeneous remote sensing images using an image translation network to obtain translated images; respectively obtaining the wavelet-domain difference image and the spatial-domain difference image between the translated images and the target heterogeneous remote sensing images; synthesizing the fused difference image between the translated images and the target heterogeneous remote sensing images according to the wavelet-domain difference image and the spatial-domain difference image; using threshold segmentation to divide the pixel points in the fused difference image to obtain the change detection result of the target heterogeneous remote sensing images. The present invention can generate high-quality difference images, thereby helping to accurately identify the changed areas in the images and improving the accuracy of change detection.

[0005] The present invention adopts the following technical solutions. A method for detecting changes in heterogeneous remote sensing images includes: Obtaining target heterogeneous remote sensing images; the target heterogeneous remote sensing images include the first X-domain remote sensing image and the first Y-domain remote sensing image; Performing stationary wavelet decomposition on the target heterogeneous remote sensing images, and processing the decomposed target heterogeneous remote sensing images using an image translation network to obtain translated images; The translated images include: the second Y-domain translated image corresponding to the first X-domain remote sensing image; the second X-domain translated image corresponding to the first Y-domain remote sensing image; Respectively obtaining the wavelet-domain difference image and the spatial-domain difference image between the translated images and the target heterogeneous remote sensing images; Synthesizing the fused difference image between the translated images and the target heterogeneous remote sensing images according to the wavelet-domain difference image and the spatial-domain difference image; Using a threshold segmentation algorithm to divide the pixel points in the fused difference image to obtain the change detection result of the target heterogeneous remote sensing images.

[0006] Further, the image translation network includes: a first generator and a second generator; The first generator is used to convert the first X-domain remote sensing image into the second Y-domain translated image, and it includes a first encoder and a first decoder; the first encoder and the first decoder adopt skip connections; The second generator is used to convert the first Y-domain remote sensing image into the second X-domain translated image, and it includes a second encoder and a second decoder; the second encoder and the second decoder adopt skip connections.

[0007] Further, the image translation network further includes: Inputting the second Y-domain translated image into the second generator to obtain the third X-domain reconstructed image; Inputting the second X-domain translated image into the first generator to obtain the third Y-domain reconstructed image.

[0008] Furthermore, the image translation network adopts a cross - domain triple loss function, specifically: The cross - domain triple loss function includes: style loss, content loss, and reconstruction loss, expressed as: ; Wherein, represents the cross - domain triple loss function, is the style loss, is the content loss, is the reconstruction loss; The style loss is expressed as: ; Wherein, represents the style loss, represents the high - frequency component in the remote sensing image of the first X domain, represents the high - frequency component in the translated image of the second X domain, represents the high - frequency component in the remote sensing image of the first Y domain, represents the high - frequency component in the translated image of the second Y domain, represents the L1 norm; The content loss is expressed as: ; Wherein, represents the content loss, represents the pixel value of the invariant region in the target heterologous remote sensing image, represents the low - frequency component in the remote sensing image of the first X domain, represents the low - frequency component in the translated image of the second X domain, represents the low - frequency component in the remote sensing image of the first Y domain, represents the low - frequency component in the translated image of the second Y domain, represents the L1 norm; The reconstruction loss is expressed as: ; Wherein, represents the reconstruction loss, represents the feature component in the remote sensing image of the first X domain, represents the feature component in the reconstructed image of the third X domain, represents the feature component in the remote sensing image of the first Y domain, represents the feature component in the reconstructed image of the third Y domain, represents the L1 norm.

[0009] Furthermore, the acquisition method of the wavelet - domain difference image is: Obtain the first low-frequency component in the first X-domain remote sensing image and the second low-frequency component in the second X-domain translated image; Obtain the first X-domain difference image according to the Euclidean distance between the first low-frequency component and the second low-frequency component; Obtain the third low-frequency component in the first Y-domain remote sensing image and the fourth low-frequency component in the second Y-domain translated image; Obtain the first Y-domain difference image according to the Euclidean distance between the third low-frequency component and the fourth low-frequency component; Fuse the first X-domain difference image and the first Y-domain difference image by using an adaptive weight fusion method to obtain a wavelet-domain difference image.

[0010] Furthermore, the acquisition method of the spatial-domain difference image is as follows: Input the first X-domain remote sensing image and the second X-domain translated image into the first encoder to obtain the first feature vector corresponding to the first X-domain remote sensing image and the second feature vector corresponding to the second X-domain translated image; Obtain the second X-domain difference image according to the Euclidean distance between the first feature vector and the second feature vector; Input the first Y-domain remote sensing image and the second Y-domain translated image into the second encoder to obtain the third feature vector corresponding to the first Y-domain remote sensing image and the fourth feature vector corresponding to the second Y-domain translated image; Obtain the second Y-domain difference image according to the Euclidean distance between the third feature vector and the fourth feature vector; Fuse the second X-domain difference image and the second Y-domain difference image by using an adaptive weight fusion method to obtain a spatial-domain difference image.

[0011] Furthermore, the synthesis method of the fused difference image is as follows: Fuse the wavelet-domain difference image and the spatial-domain difference image by using an information entropy weighted fusion method to obtain a fused difference image.

[0012] Based on the proposed method for detecting changes in heterogeneous remote sensing images, the present invention also proposes a system for detecting changes in heterogeneous remote sensing images, which includes: An image acquisition module for acquiring target heterogeneous remote sensing images; An image translation module for performing stationary wavelet decomposition on the target heterogeneous remote sensing images and processing the decomposed target heterogeneous remote sensing images by using an image translation network to obtain translated images; A difference image extraction module for respectively obtaining the wavelet-domain difference image and the spatial-domain difference image between the translated images and the target heterogeneous remote sensing images; The differential image fusion module is used to synthesize the fusion differential image between the translated image and the target heterologous remote sensing image according to the wavelet-domain differential image and the spatial-domain differential image; The image detection module is used to divide the pixel points in the fusion differential image by using a threshold segmentation algorithm to obtain the change detection result of the target heterologous remote sensing image.

[0013] The beneficial effects of the present invention are as follows: The image translation network proposed by the present invention only includes two generators with the same structure and no discriminator, which is simpler and has fewer parameters compared with other network structures. By adding skip connections, the feature extraction ability of the network is further enhanced to help generate clearer differential images in the subsequent process; At the same time, the image translation network is optimized by using a simple and effective cross-domain triple loss function. The style loss aims to narrow the style difference between the translated image and the target image, the content loss aims to amplify the content difference in the changed area between the translated image and the target image, and the reconstruction loss is used to prevent the loss of image information, so as to better highlight the changes between heterologous images; For the generated differential images, the present invention adopts an adaptive weight fusion strategy and an information entropy weighted fusion strategy to fuse the forward and backward differential images in the wavelet domain and the spatial domain, which can fully extract the changes in the translation process, thereby generating higher-quality differential maps, which further helps to improve the performance of change detection in image change detection. Description of the Drawings

[0014] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained according to these drawings.

[0015] Figure 1 It is a schematic flow chart of a method for detecting changes in heterologous remote sensing images according to an embodiment of the present invention; Figure 2 It is a schematic structural diagram of an encoder-decoder skip connection method according to an embodiment of the present invention; Figure 3 It is a schematic diagram of the translation process of an image translation network according to an embodiment of the present invention; Figure 4 It is a schematic diagram for comparing the detection results of different detection methods according to an embodiment of the present invention; where, Figure 4 (a) is the pre-event image, Figure 4 (b) is the post-event image, Figure 4 (c) is the actual ground change map, Figure 4 (d) is a schematic diagram of the change detection result using the X-Net network, Figure 4(e) Schematic diagram of the change detection result using the ACE-Net network, Figure 4 (f) Schematic diagram of the change detection result using the GIR-MRF method, Figure 4 (g) Schematic diagram of the detection result using the CAAE method, Figure 4 (h) Schematic diagram of the detection result using the AGSCC method, Figure 4 (i) Schematic diagram of the detection result using the SDIR method, Figure 4 (j) Schematic diagram of the detection result of the method proposed by the present invention; Figure 5 It is a schematic diagram of the structure of a heterogeneous remote sensing image change detection system according to an embodiment of the present invention. Specific implementation manner

[0016] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0017] A schematic diagram of the process of a heterogeneous remote sensing image change detection method according to an embodiment of the present invention is as Figure 1 shown and includes: Obtain the target heterogeneous remote sensing image; Heterogeneous remote sensing images refer to remote sensing images with diverse data characteristics obtained through different sensors, different platforms, or different times. Their data differences are reflected in aspects such as spectral resolution, spatial resolution, temporal resolution, or imaging principles (such as optical, radar, infrared, etc.). In the embodiments of the present invention, the target heterogeneous remote sensing images include the first X-domain remote sensing image and the first Y-domain remote sensing image; correspondingly, the first X-domain remote sensing image in the embodiments of the present invention can be regarded as the pre-change image, and the first Y-domain remote sensing image can be regarded as the post-change image. The pre-change image and the post-change image are in different image domains, and different image domains are remote sensing images obtained through different methods. In the embodiments of the present invention, it can be obtained through any one of the methods such as the cooperation of multi-sensors and satellites, the fusion of UAV images and satellite images, and the combination of visible light images and thermal infrared images. When training the image translation network of the present invention, a large number of remote sensing images are downloaded by accessing the existing remote sensing image library to facilitate the construction of a remote sensing image dataset, thereby saving training time.

[0018] Perform stationary wavelet decomposition on the target heterogeneous remote sensing image, and use the image translation network to process the decomposed target heterogeneous remote sensing image to obtain a translated image; In the embodiments of the present invention, the stationary wavelet decomposition method avoids the downsampling operation, making the decomposition process translation invariant, and can better retain the local features and detail information of the image. By performing stationary wavelet decomposition on the target heterogeneous remote sensing image, the target heterogeneous remote sensing image is decomposed into a low-frequency component and three high-frequency detail components. Among them, the low-frequency component represents the approximate part of the image, and the three high-frequency components respectively represent the horizontal, vertical, and diagonal direction parts of the image. Thus, the low-frequency component can represent the content information of the target heterogeneous remote sensing image, and the high-frequency component can represent the style information of the image.

[0019] In the embodiments of the present invention, the image translation network includes a first generator and a second generator, which form a pseudo-conjoined network through the first generator and the second generator, and the first generator and the second generator share the same structure; The first generator includes a first encoder and a first decoder, which are used to convert the first X-domain remote sensing image into a second Y-domain translated image; the second generator includes a second encoder and a second decoder, which are used to convert the first Y-domain remote sensing image into a second X-domain translated image; skip connections are adopted between the first encoder and the first decoder, and between the second encoder and the second decoder, so as to further enhance the feature representation ability of the generator.

[0020] The skip connection method is also called the residual connection. As Figure 2 shown, in a first encoder-decoder structure in the embodiments of the present invention, the first encoder is responsible for converting the input sequence into an intermediate representation (feature vector), and the first decoder generates an output sequence based on this intermediate representation. That is, the skip connection between the first encoder and the first decoder is to directly transfer some output information of the first encoder to certain layers of the first decoder without passing through all the intermediate layers. At the same time, in the first encoder-decoder structure in the embodiments of the present invention, the convolution type and parameters of each layer are shown in Table 1, and the same applies to the second encoder and the second decoder, and the convolution type and parameters are the same as those of the first encoder-decoder structure.

[0021] Table 1 Schematic table of convolution type and parameters of each layer in the encoder-decoder structure ; In a specific embodiment of the present invention: The translated images include: the second Y-domain translated image corresponding to the first X-domain remote sensing image; the second X-domain translated image corresponding to the first Y-domain remote sensing image; the translated images have the same style as the target heterogeneous remote sensing image but different contents. In order to better optimize the image translation network, the present invention adopts a cross-domain triple loss function to reduce the style difference between the translated image and the target image, and amplify the content difference in the changing area between the translated image and the target image, so as to better highlight the changes between the images. The translation process is asFigure 3 As shown; in the embodiments of the present invention, the cross-domain triple loss function is established based on style loss, content loss, and reconstruction loss, and is expressed as: ; Wherein, represents the cross-domain triple loss function, is the style loss, is the content loss, is the reconstruction loss; The style loss is expressed as: ; Wherein, represents the style loss, represents the high-frequency component in the first X-domain remote sensing image, represents the high-frequency component in the second X-domain translated image, represents the high-frequency component in the first Y-domain remote sensing image, represents the high-frequency component in the second Y-domain translated image, represents the L1 norm; The content loss is expressed as: ; Wherein, represents the content loss, represents the pixel value of the invariant region in the target heterologous remote sensing image, represents the low-frequency component in the first X-domain remote sensing image, represents the low-frequency component in the second X-domain translated image, represents the low-frequency component in the first Y-domain remote sensing image, represents the low-frequency component in the second Y-domain translated image, represents the L1 norm; The reconstruction loss is expressed as: ; Wherein, represents the reconstruction loss, represents the feature component in the first X-domain remote sensing image, represents the feature component in the third X-domain reconstructed image, represents the feature component in the first Y-domain remote sensing image, represents the feature component in the third Y-domain reconstructed image, represents the L1 norm.

[0022] In a specific embodiment of the present invention, when training the image translation network, in addition to using the downloaded remote sensing image dataset, it is also necessary to obtain the label mask matrix corresponding to the remote sensing image dataset, that is, each remote sensing image has a corresponding label mask matrix. Each element in the label mask matrix corresponds to a pixel in the remote sensing image. Since the label mask matrix is randomly initialized, the invariant regions in the remote sensing image dataset obtained according to the label mask matrix are not completely correct. Therefore, after completing the training of the image translation network, the label mask matrix can be updated by translating the image to obtain better change detection results. On this basis, the pixel values of the invariant regions in the target heterologous remote sensing image can be obtained by the following formula: ; wherein, is the pixel value of the invariant region in the target heterologous remote sensing image, is the label mask matrix; when the pixel value of the invariant region in the target heterologous remote sensing image is obtained, image translation can be performed, and then the differential image can be obtained to generate the corresponding binary change map. As the number of iterative updates increases, the accuracy of the obtained binary change map will gradually improve.

[0023] The wavelet domain difference image and the spatial domain difference image between the translated image and the target heterologous remote sensing image are obtained respectively; In the embodiment of the present invention, the method for obtaining the wavelet domain difference image between the target heterologous remote sensing image and the translated image is as follows: obtaining the first low-frequency component in the first X-domain remote sensing image and the second low-frequency component in the second X-domain translated image; obtaining the first X-domain difference image according to the Euclidean distance between the first low-frequency component and the second low-frequency component; obtaining the third low-frequency component in the first Y-domain remote sensing image and the fourth low-frequency component in the second Y-domain translated image; obtaining the first Y-domain difference image according to the Euclidean distance between the third low-frequency component and the fourth low-frequency component; and fusing the first X-domain difference image and the first Y-domain difference image by using the adaptive weight fusion method to obtain the wavelet domain difference image between the target heterologous remote sensing image and the translated image.

[0024] The method for obtaining the spatial domain difference image between the translated image and the target heterogeneous remote sensing image is as follows: Input the first X-domain remote sensing image and the second X-domain translated image into the first encoder to obtain the first feature vector corresponding to the first X-domain remote sensing image and the second feature vector corresponding to the second X-domain translated image; obtain the second X-domain difference image according to the Euclidean distance between the first feature vector and the second feature vector; input the first Y-domain remote sensing image and the second Y-domain translated image into the second encoder to obtain the third feature vector corresponding to the first Y-domain remote sensing image and the fourth feature vector corresponding to the second Y-domain translated image; obtain the second Y-domain difference image according to the Euclidean distance between the third feature vector and the fourth feature vector; use the adaptive weight fusion method to fuse the second X-domain difference image and the second Y-domain difference image to obtain the spatial domain difference image.

[0025] In a specific embodiment of the present invention: The fusion of the first X-domain difference image and the first Y-domain difference image using the adaptive weight fusion method can be expressed as: ; where represents the wavelet domain difference image, represents the first X-domain difference image, represents the first Y-domain difference image, represents the standard deviation of the first X-domain difference image, represents the standard deviation of the first Y-domain difference image.

[0026] In a specific embodiment of the present invention: The fusion of the second X-domain difference image and the second Y-domain difference image using the adaptive weight fusion method to obtain the spatial domain difference image can be expressed as: ; where represents the spatial domain difference image, represents the second X-domain difference image, represents the second Y-domain difference image, represents the standard deviation of the second X-domain difference image, represents the standard deviation of the second Y-domain difference image.

[0027] Synthesize the fusion difference image between the translated image and the target heterogeneous remote sensing image according to the wavelet domain difference image and the spatial domain difference image; In the embodiment of the present invention, the wavelet domain difference image and the spatial domain difference image are synthesized using the information entropy weighting method to obtain the fusion difference image between the translated image and the target heterogeneous remote sensing image, and the synthesis process can be expressed as: ; Among them, represents the fusion difference image between the translated image and the target heterogeneous remote sensing image, represents the wavelet domain difference image, represents the information entropy of the wavelet domain difference image, represents the spatial domain difference image, represents the information entropy of the spatial domain difference image.

[0028] In a specific embodiment of the present invention: The wavelet domain information contains the multi-scale frequency characteristics of the image, can perform multi-scale decomposition of the image from the frequency domain perspective, effectively capture the change details of the image at different frequencies, extract the detail and texture features of the image, and is especially good at capturing the local change information of the image; the spatial domain information has the original spatial structure and texture information of the image, intuitively reflects the pixel gray distribution and spatial structure relationship of the image, and retains the original morphological features of the image. The two complement each other's advantages and provide a rich data basis for the construction of the difference map. However, how to reasonably allocate the weights of the two in the fusion process of the wavelet domain difference image and the spatial domain difference image becomes the key to improving the fusion effect. In this regard, a deep neural network model is constructed in the embodiment of the present invention. The obtained wavelet domain difference image and spatial domain difference image are respectively used as the inputs of the network. Through the complex operations of multiple layers of neurons, the importance of the two types of information for the construction of the difference image in different scenarios is learned, and the optimal information entropy weight value is determined.

[0029] Specifically, when training the deep neural network model, first preprocess the obtained wavelet domain difference image and spatial domain difference image, so as to extract the wavelet domain coefficient matrix in the wavelet domain difference image and the spatial domain feature map in the spatial domain difference image respectively; further, the deep neural network model constructed in the embodiment of the present invention adopts a residual network combined with an attention mechanism. The input layer of the network model receives the processed wavelet domain coefficient matrix and spatial domain feature map, and respectively performs feature extraction on the two types of information through parallel convolutional blocks. Each convolutional block includes a 3×3 convolutional layer, a batch normalization layer, and a ReLU activation function, which enhances the non-linear expression ability of the network while suppressing the vanishing gradient.

[0030] In the feature fusion stage of the deep neural network model, the embodiment of the present invention introduces a spatial attention module and a channel attention module. The spatial attention module calculates the spatial correlation between pixel points and highlights the spatial position information of the changed area; the channel attention module strengthens the key information channels and suppresses redundant information according to the feature response intensity between channels. For example, when processing post-earthquake disaster images, the spatial attention module can focus on the spatial distribution area of building collapses, and the channel attention module enhances the feature channels reflecting structural changes, enabling the network to more accurately capture disaster traces.

[0031] To further improve the generalization ability of the deep neural network model, in the embodiments of the present invention, a global average pooling layer and a fully connected layer are introduced at the end of the network. The fused features are mapped into an adaptive weight vector, which is normalized by the Softmax function and then weighted and fused with the wavelet domain coefficient matrix and the spatial domain feature map to ensure that the sum of the weights is 1, realizing the effective fusion of the two types of information.

[0032] The training of the deep neural network model adopts an end-to-end supervised learning method. To obtain the optimal difference image, the mean square error (MSE) and the structural similarity index (SSIM) are set as the joint loss function. The Adam algorithm is selected as the optimizer, the initial learning rate is set to 0.001, and the cosine annealing learning rate adjustment strategy is adopted to avoid falling into local optima while ensuring the convergence speed.

[0033] The information entropy weighted fusion strategy optimized by the neural network in the embodiments of the present invention can dynamically and accurately fuse the wavelet domain information and the spatial domain information. The fused difference image obtained through this strategy can not only clearly present the changed areas, effectively distinguish the changed class and the unchanged class, but also significantly reduce the influence of the background and noise, having broad application prospects and great technical advantages in the field of remote sensing image change detection.

[0034] The threshold segmentation algorithm is used to divide the pixel points in the fused difference image to obtain the change detection result of the target heterologous remote sensing image.

[0035] In the embodiments of the present invention, the threshold segmentation algorithm is used to divide the pixel points in the fused difference image. The image pixels are divided into two categories (foreground and background). By calculating the between-class variance of the two types of pixels under different thresholds, the threshold that maximizes the between-class variance is selected as the optimal threshold. Usually, the optimal threshold can be determined by the threshold segmentation algorithm itself, so as to distinguish the changed area and the unchanged area in the fused difference image, and mark the pixel points at the changed positions and the unchanged positions in the fused difference image as "1" and "0" respectively, thus realizing the detection of the deformation of the heterologous remote sensing image.

[0036] In another specific embodiment of the present invention: Based on the heterologous remote sensing image change detection method, the present invention also proposes a heterologous remote sensing image change detection system, the structure of which is as Figure 5 shown. The system includes: An image acquisition module, used to acquire the target heterologous remote sensing image; An image translation module, used to perform stationary wavelet decomposition on the target heterologous remote sensing image and process the decomposed target heterologous remote sensing image by using an image translation network to obtain a translated image; The difference image extraction module is used to respectively obtain the wavelet-domain difference image and the spatial-domain difference image between the translated image and the target heterogeneous remote sensing image; The difference image fusion module is used to synthesize the fusion difference image between the translated image and the target heterogeneous remote sensing image according to the wavelet-domain difference image and the spatial-domain difference image; The image detection module is used to divide the pixel points in the fusion difference image by using a threshold segmentation algorithm to obtain the change detection result of the target heterogeneous remote sensing image.

[0037] In an experimental embodiment of the present invention: The detection method proposed by the present invention is experimentally compared with several existing methods. The schematic diagram of the comparison results is as Figure 4 shown. In the embodiment of the present invention, three datasets for testing are selected to perform image change detection respectively, and the false alarm, missed alarm, overall error, overall accuracy, F1 value and Kappa coefficient of the detection results are calculated respectively to evaluate the effects of different detection methods. The quantization results are shown in Table 2, Table 3 and Table 4 respectively: Table 2 Schematic table of evaluation results on the Sardinia dataset ; Table 3 Schematic table of evaluation results on the Gloucester dataset ; Table 4 Schematic table of evaluation results on the California dataset ; Figure 4 In Figure 4 (a) is the pre-event image, Figure 4 (b) is the post-event image, Figure 4 (c) is the actual ground change map, Figure 4 (d) is the change detection result using the X-Net network, Figure 4 (e) is the change detection result using the ACE-Net network, Figure 4 (f) is the change detection result using the GIR-MRF method, Figure 4 (g) is the detection result using the CAAE method, Figure 4 (h) is the detection result using the AGSCC method, Figure 4 (i) is the detection result using the SDIR method, Figure 4 (j) is the detection result of the method proposed by the present invention, and Figure 4 from top to bottom in , the first row is the Sardinia (Sardinia, Italy) dataset, the second row is the Gloucester dataset, and the third row is the California dataset,Figure 4 The changed regions and unchanged regions are represented by white and black respectively, and red represents FP, that is, the part that the model misjudges as a changed region but is actually unchanged; green represents FN, that is, the part that the model misjudges as an unchanged region but actually has changed. It can be seen that the change map generated by the method proposed in the present invention is relatively clear, and the highest F1 value and Kappa coefficient are obtained on all three datasets tested, thus verifying the effectiveness and practicability of the heterologous image change detection method proposed in the present invention.

[0038] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for detecting changes in heterogeneous remote sensing images, characterized in that, Including: Obtain a target heterogeneous remote sensing image; The target heterogeneous remote sensing image includes a first X-domain remote sensing image and a first Y-domain remote sensing image; Perform stationary wavelet decomposition on the target heterogeneous remote sensing image, and use an image translation network to process the decomposed target heterogeneous remote sensing image to obtain a translated image; The translated image includes: a second Y-domain translated image corresponding to the first X-domain remote sensing image; a second X-domain translated image corresponding to the first Y-domain remote sensing image; The image translation network includes: a first generator and a second generator; The first generator is used to convert the first X-domain remote sensing image into a second Y-domain translated image, and it includes a first encoder and a first decoder; the first encoder and the first decoder use skip connections; The second generator is used to convert the first Y-domain remote sensing image into a second X-domain translated image, and it includes a second encoder and a second decoder; the second encoder and the second decoder use skip connections; Respectively obtain the wavelet-domain difference image and the spatial-domain difference image between the translated image and the target heterogeneous remote sensing image; According to the wavelet-domain difference image and the spatial-domain difference image, synthesize the fusion difference image between the translated image and the target heterogeneous remote sensing image; Use a threshold segmentation algorithm to divide the pixel points in the fusion difference image to obtain the change detection result of the target heterogeneous remote sensing image.

2. The method for detecting changes in heterogeneous remote sensing images according to claim 1, wherein: The image translation network further includes: Input the second Y-domain translated image into the second generator to obtain a third X-domain reconstructed image; Input the second X-domain translated image into the first generator to obtain a third Y-domain reconstructed image.

3. A method for detecting changes in heterogeneous remote sensing images according to claim 2, characterized in that: The image translation network uses a cross-domain triple loss function, specifically: The cross-domain triple loss function includes: style loss, content loss, and reconstruction loss, expressed as: ； Among them, represents a cross-domain triple loss function, is the style loss, is the content loss, is the reconstruction loss; The style loss is expressed as: ； Among them, represents the style loss, represents the high-frequency component in the first X-domain remote sensing image, represents the high-frequency component in the second X-domain translated image, represents the high-frequency component in the first Y-domain remote sensing image, represents the high-frequency component in the second Y-domain translated image, represents the L1 norm; The content loss is expressed as: ； Among them, represents content loss, represents the pixel value of the invariant region in the target heterologous remote sensing image, represents the low-frequency component in the first X-domain remote sensing image, represents the low-frequency component in the second X-domain translated image, represents the low-frequency component in the first Y-domain remote sensing image, represents the low-frequency component in the second Y-domain translated image, represents the L1 norm; The reconstruction loss is expressed as: ； Among them, represents the reconstruction loss, represents the feature component in the first X-domain remote sensing image, represents the feature component in the third X-domain reconstructed image, represents the feature component in the first Y-domain remote sensing image, represents the feature component in the third Y-domain reconstructed image, represents the L1 norm.

4. A method for detecting changes in heterogeneous remote sensing images according to claim 1, characterized in that: The acquisition method of the wavelet-domain difference image is: Obtain the first low-frequency component in the first X-domain remote sensing image and the second low-frequency component in the second X-domain translated image; According to the Euclidean distance between the first low-frequency component and the second low-frequency component, obtain the first X-domain difference image; Obtain the third low-frequency component in the first Y-domain remote sensing image and the fourth low-frequency component in the second Y-domain translated image; According to the Euclidean distance between the third low-frequency component and the fourth low-frequency component, obtain the first Y-domain difference image; Use an adaptive weight fusion method to fuse the first X-domain difference image and the first Y-domain difference image to obtain the wavelet-domain difference image.

5. A method for detecting changes in heterogeneous remote sensing images according to claim 2, characterized in that: The acquisition method of the spatial-domain difference image is: Input the first X-domain remote sensing image and the second X-domain translated image into the first encoder to obtain the first feature vector corresponding to the first X-domain remote sensing image and the second feature vector corresponding to the second X-domain translated image; According to the Euclidean distance between the first feature vector and the second feature vector, obtain the second X-domain difference image; Input the first Y-domain remote sensing image and the second Y-domain translated image into the second encoder to obtain the third feature vector corresponding to the first Y-domain remote sensing image and the fourth feature vector corresponding to the second Y-domain translated image; Obtain the second Y-domain difference image according to the Euclidean distance between the third feature vector and the fourth feature vector; Fuse the second X-domain difference image and the second Y-domain difference image by using an adaptive weight fusion method to obtain the spatial-domain difference image.

6. The method for detecting changes in heterogeneous remote sensing images according to claim 1, characterized in that: The synthesis method of the fused difference image is as follows: Use the information entropy weighted fusion method to fuse the wavelet-domain difference image and the spatial-domain difference image to obtain the fused difference image.

7. A heterologous remote sensing image change detection system, characterized in that, It includes: An image acquisition module for acquiring target heterogeneous remote sensing images; An image translation module for performing stationary wavelet decomposition on the target heterogeneous remote sensing image and processing the decomposed target heterogeneous remote sensing image by using an image translation network to obtain a translated image; A difference image extraction module for respectively obtaining the wavelet-domain difference image and the spatial-domain difference image between the translated image and the target heterogeneous remote sensing image; A difference image fusion module for synthesizing the fused difference image between the translated image and the target heterogeneous remote sensing image according to the wavelet-domain difference image and the spatial-domain difference image; An image detection module for dividing the pixel points in the fused difference image by using a threshold segmentation algorithm to obtain the change detection result of the target heterogeneous remote sensing image.

Images