A change detection method and device based on high-quality feature selection

Abstract

The application discloses a change detection method and device based on high-quality feature selection, which comprises the following steps: feature extraction is performed on an input image pair through a twin U-shaped network to obtain image features of different scales, and the image features are used as input features to construct a feature pool; the twin U-shaped network has full attention enhancement characteristics and is based on a VGG16bn network; feature selection is performed according to the contribution degree of different change-related features in the feature pool to obtain high-quality change features; pyramid feature fusion is performed on the change features of different scales in combination with attention; a convolution operation is added to each layer of the pyramid feature fusion output, each layer generates a change mask, and five predicted change masks are spliced and subjected to a convolution operation to obtain a final predicted change mask; a loss function is constructed based on the attention pyramid fusion and the final predicted change mask, and the changed image region is detected through the loss function. The device comprises a processor and a memory.

Description

Technical Field

[0001] This invention relates to the field of change detection, and in particular to a change detection method and apparatus based on high-quality feature selection. Background Technology

[0002] Change detection aims to detect whether a target region has changed, and can distinguish between real changes and changes in the background environment between pairs of images taken at different times from the same scene. [1] It has wide applications in resource monitoring, urban management, damage assessment, autonomous driving, and military applications. [2] .

[0003] Existing deep learning-based change detection methods can identify changes from a large number of training images and generate relatively accurate change detection results. [3,4,5] However, extracting features relevant to change is the main challenge in generating high-quality change detection results. Previous work [5,6] Two images were pre-combined into six channels before being fed into a fully convolutional network for change detection. However, this approach failed to better explore feature information to distinguish between changing regions and the background, thus limiting the change detection results. Later work... [7,8,9] These methods tend to extract features from each image using Siamese networks and then combine features with differences or cascades. While change detectors using the latter combination can achieve better change detection results, they ignore the complementarity of different types of features. For example, the FCNCD method.

[10] The ADCDNet method uses channel feature concatenation to obtain variation features. [8] The former approach uses absolute differences to obtain change features. However, while it preserves all information from the image pair, it correspondingly weakens the ability to represent changes, making it difficult to distinguish between background and foreground, leading to false detections of background changes. The latter approach emphasizes change information between image pairs to some extent, but ignores the semantic information of the changing object, easily leading to the loss of edge feature information in the detected image and incomplete target information.

[0004] Therefore, the key to obtaining high-quality change detection results lies in obtaining high-quality change features.

[0005] References

[0006] [1]Singh, Ashbindu. "Review article digital change detection techniques using remotely-sensed data." International journal of remote sensing 10.6(1989):989-1003.

[0007] [2] Sui Haigang, Feng Wenqing, Sun Kaimin, Xu Chuan, A review of methods for detecting changes in multi-temporal remote sensing images. vol.43.no.12,2018.

[0008] [3]Daudt, Rodrigo Caye, Bertr Le Saux, and Alexandre Boulch. "Fullyconvolutional siamese networks for change detection." 2018 25th IEEEInternational Conference on Image Processing (ICIP). IEEE, 2018.

[0009] [4] Zhang, Chenxiao, et al. "A deeply supervised image fusion network for change detection in high resolution bi-temporal remote sensing images." ISPRSJournal of Photogrammetry and Remote Sensing 166(2020):183-200.

[0010] [5] Alcantarilla, Pablo F., et al. "Street-view change detection with deconvolutional networks." Autonomous Robots 42.7 (2018): 1301-1322.

[0011] [6]Sakurada,Ken,Mikiya Shibuya,and Weimin Wang."Weakly supervisedsilhouette-based semantic scene change detection."2020IEEE Internationalconference on robotics and automation(ICRA).IEEE,2020.

[0012] [7]Chen,Hao,Zipeng Qi,and Zhenwei Shi."Remote sensing image changedetection with transformers."IEEE Transactions on Geoscience and RemoteSensing 60(2021):1-14.

[0013] [8]Huang,Rui,et al."Change detection with absolute difference ofmultiscale deep features."Neurocomputing 418(2020):102-113.

[0014] [9]Xu,Jialang,et al."An Adaptive Multi-Scale and Multi-Level FeaturesFusion Network with Perceptual Loss for Change Detection."ICASSP 2021-2021IEEE International Conference on Acoustics,Speech and Signal Processing(ICASSP).IEEE,2021.

[0015]

[10] Long,Jonathan,Evan Shelhamer,and Trevor Darrell."Fullyconvolutional networks for semantic segmentation."Proceedings of the IEEEconference on computer vision and pattern recognition.2015. Summary of the Invention

[0016] This invention provides a change detection method and apparatus based on high-quality feature selection. The invention uses a VGG16bn-based Siamese U-shaped network with full attention enhancement to extract five different scale features from input image pairs. These features are then used as input features to construct a feature pool. High-quality change features are obtained by selecting features based on the contribution of different change-related features in the feature pool. Finally, pyramid-style feature fusion is performed on the change features at different scales using attention to address the problem of incomplete or indistinct change regions. Details are described below.

[0017] A change detection method based on high-quality feature selection, the method comprising:

[0018] The input image pairs are feature extracted using a Siamese U-shaped network to obtain image features at different scales, and these image features are used as input features to construct a feature pool. The Siamese U-shaped network has full attention enhancement characteristics and is based on the VGG16bn network.

[0019] Feature selection is performed based on the contribution level of different change-related features in the feature pool to obtain high-quality change features; pyramid-style feature fusion is performed on change features at different scales by combining attention.

[0020] Convolution operations are added to each layer of features output by pyramid feature fusion, and a change mask is generated for each layer. The five predicted change masks are concatenated and convolved to obtain the final predicted change mask.

[0021] A loss function is constructed based on attention pyramid fusion and the final predicted change mask, and the changed image regions are detected through the loss function.

[0022] Specifically, constructing a feature pool by using image features as input features involves:

[0023] The paired features obtained from the Siamese network are used as input, and the concatenated features are obtained through concatenation and convolution operations. The absolute difference features are obtained through feature subtraction and convolution operations. The extracted paired features, concatenated features, and absolute difference features are used to construct a feature pool.

[0024] Furthermore, the step of selecting features based on the contribution of input features to obtain high-quality changing features specifically involves: treating these three different types of features as expert features, randomly initializing an updatable weight matrix, selecting features from the feature pool, and summing the weights based on their importance to obtain high-quality changing features.

[0025] The high-quality variation characteristics are as follows:

[0026]

[0027]

[0028] Among them, E k (,:,:,j) represents the feature of the j-th channel of the k-th expert, and W is a learnable parameter. This represents the weight value of the k-th expert feature in the j-th channel after normalization. Let represent the weighted value of the k-th expert feature in the j-th channel, and m represent the m-th feature among the k expert features. This represents the summation of the weighted transformation values ​​of the k expert features in the j-th channel.

[0029] Furthermore, the change mask is:

[0030]

[0031]

[0032] Among them, M i This represents the change mask for each layer of prediction. This indicates the use of features prior to prediction at each layer. Guided by features processed by the attention pyramid fusion module The features generated for this layer are used to predict the change mask for this layer. Conv1(·,1) indicates that the layer is processed by a convolution with a 1×1 kernel. This represents the features after processing by the attention pyramid fusion module. After convolution, the deepest layer of features is generated to predict the change mask.

[0033] A second aspect is a change detection device based on high-quality feature selection, the device comprising: a processor and a memory, the memory storing program instructions, the processor calling the program instructions stored in the memory to cause the device to perform the method steps of any one of the first aspects.

[0034] Third aspect, a computer-readable storage medium storing a computer program, the computer program including program instructions that, when executed by a processor, cause the processor to perform the steps of the method described in any of the first aspects.

[0035] The beneficial effects of the technical solution provided by this invention are:

[0036] 1. This invention uses image pairs acquired at different times as input and utilizes a Siamese network with full attention enhancement to extract multi-scale features of the image. The introduction of the attention mechanism makes the extracted features more complete and focused on the foreground target, reduces the interference of invalid background information on the detection results, and makes the prediction results closer to the real change results.

[0037] 2. The high-quality feature selection network constructed in this invention uses feature pairs extracted by the Siamese U-shaped network as input to construct a feature pool of change-related features. Then, feature selection and combination are performed based on the contribution degree of different change-related features in the feature pool to obtain high-quality change features. This solves the problem of edge incompleteness or lack of prominence of change areas caused by relying on a single feature in the past.

[0038] 3. Based on the fact that changed images contain richer feature information, this invention adopts an attention pyramid fusion method when performing multi-scale feature fusion. This not only effectively solves the information differences that exist when fusing features at different levels, but also allows the changed regions in the fused features to receive more attention. At the same time, it generates preliminary prediction maps at multiple levels and fuses them to generate the final prediction map. This helps to utilize high-level semantic features and low-level texture features to complement and promote each other, and solves the problem of incomplete changed regions in the prediction map.

[0039] 4. Experimental results on VL-CMU (Visual Localization CMU dataset), PCD (Panorama Change Detection dataset), and CDnet2014 (Change Detection Video Sequence dataset) datasets show that the method has a significant gain effect. Attached Figure Description

[0040] Figure 1 This is a schematic diagram of the change detection network proposed in this invention;

[0041] Figure 2 This is a schematic diagram of the structure of the high-quality feature selection network proposed in this invention;

[0042] Figure 3 This is a schematic diagram comparing the experimental results of the change detection method proposed in this invention with other existing change detection methods;

[0043] Figure 4 A flowchart of a change detection method based on high-quality feature selection;

[0044] Figure 5 This is a schematic diagram of a change detection device based on high-quality feature selection. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below.

[0046] This method proposes a change detection method and device based on high-quality feature selection to address the problem of image change region detection. It can detect changing objects in the image to be detected and more closely approximate the actual changes.

[0047] Example 1

[0048] A change detection method based on high-quality feature selection, see [link to relevant documentation]. Figures 1 to 3 The method includes the following steps:

[0049] I. Constructing a Siamese U-shaped network with full attention enhancement

[0050] In this embodiment of the invention, a twin U-shaped network with full attention enhancement is constructed as a feature extraction network. The VGG16bn network is used for basic feature extraction, and then the full attention enhancement module is used to enhance the deep features. Then, the basic features are enhanced in a top-down manner to form a twin U-shaped network to extract features from paired images, resulting in five sets of features at different scales.

[0051] The network uses the image to be detected as a pair<I1,I2> As input, five sets of paired features F at different scales are used. i and (i = 1, 2, 3, 4, 5) are output.

[0052] First, the images I1 and I2 to be detected are processed through the VGG16bn network for basic feature extraction. The first layer features of the I1 image are denoted as F1. Conv1 The first layer of features in the I2 image is denoted as F2. Conv1 ; F1 Conv1 With F2 Conv1 The images are transmitted together to the second convolutional layer. The second convolutional layer transforms the 64*160*160 tensor into a 128*80*80 tensor. The third convolutional layer transforms the 128*80*80 tensor into a 256*40*40 tensor. The fourth convolutional layer transforms the 256*40*40 tensor into a 512*20*20 tensor. The fifth convolutional layer transforms the 512*20*20 tensor into a 512*10*10 tensor. After basic feature extraction, the five sets of features obtained from the paired images are represented as follows: and Where i = 1, 2, 3, 4, 5.

[0053] The full attention enhancement module applies full attention to the fifth and fourth layers of paired features to capture global information dependencies and enhance global features. Then, it sequentially concatenates and convolves the five tensors of different scales in a top-down manner to obtain five feature tensors of different scales.

[0054]

[0055]

[0056]

[0057] Where, ψ FLA (·) represents a full attention operation, i.e., attention calculation is performed simultaneously in the channel and spatial dimensions; [·] represents feature concatenation along the channel dimension; Up(·) represents feature upsampling; and Conv2(·,3) represent two 3×3 convolutional layers. Thus, we obtain the enhanced feature pairs F at five different scales for the input image pair. i It includes F1 i and

[0058] II. Constructing a High-Quality Feature Selection Network

[0059] A high-quality feature selection network is constructed, comprising modules for processing high-quality feature selection at five different scales. Each high-quality feature selection module consists of two parts: feature pooling and feature selection. Paired features obtained from the Siamese network are used as input, and concatenated features are obtained through concatenation and convolution operations. The absolute difference feature is obtained by subtracting the features, taking the absolute value, and performing a convolution operation. Using the extracted pairwise features F1 i and splicing features Absolute difference characteristics Construct a feature pool.

[0060]

[0061]

[0062] Here, abs(·) represents the absolute value operation, [·] represents feature concatenation along the channel dimension, and Conv1(·,3) and Conv1(·,1) represent 3*3 and 1*1 convolutional layers, respectively. Thus, the feature pool can be obtained.

[0063] Then, drawing on the idea of ​​hybrid experts, these three different types of features are treated as expert features, and an updatable weight matrix is ​​randomly initialized. (Where N is the number of expert features and M is the number of channels for expert features.) Feature selection is performed on the features in the feature pool, and weights are summed according to their importance to obtain high-quality variation features. The formula is as follows:

[0064]

[0065]

[0066] Among them, E k (,:,:,j) represents the feature of the j-th channel of the k-th expert, and W is a learnable parameter that is continuously updated as the network is trained. Formula (7) indicates that when generating the weight matrix, a normalization operation is first performed to convert the probability of the output value to between 0 and 1, so that the sum of the weights of the features of different experts in the same channel is 1. This represents the weight value of the k-th expert feature in the j-th channel after normalization. Let represent the weighted value of the k-th expert feature in the j-th channel, and m represent the m-th feature among the k expert features. This represents the summation of the weighted transformation values ​​of the k expert features in the j-th channel.

[0067] III. Constructing an Attention Pyramid Fusion and Variation Mask Prediction Network

[0068] Feature fusion and propagation are performed using a top-down feature fusion approach: the high-quality variation features obtained are processed by convolution and then propagated down the layers in a top-down manner. Features with smaller resolution but richer semantic information are propagated to lower layers one by one. The semantic information of the layers is combined with the detailed texture information of the lower layers. The high-level semantic information is used to guide the detailed information and refine the texture information of the variation regions of interest.

[0069] Attention pyramid fusion refers to using attention to address the information gap between features of different scales during cross-scale feature fusion, thereby improving feature fusion performance. This fusion process is a top-down process.

[0070]

[0071]

[0072] Where, Φ APF (·) indicates the attention pyramid fusion module. This represents the output features at different scales after processing by the attention pyramid fusion module. This represents the deepest output feature after processing by the attention pyramid fusion module, and Up(·) indicates that feature upsampling operation is performed.

[0073] Transformation mask prediction first adds a convolution operation to each layer of the pyramid feature fusion output, generating a transformation mask for each layer. Simultaneously, it utilizes features from the previous layer to enhance the features of the current layer, thereby achieving better prediction results.

[0074]

[0075]

[0076]

[0077] Among them, M i This represents the change mask for each layer of prediction. This indicates the use of features prior to prediction at each layer. Guided by features processed by the attention pyramid fusion module The generated feature output of this layer is used to predict the change mask in this layer. Conv1(·,1) indicates that it is processed by a convolution with a 1×1 kernel. This represents the features after processing by the attention pyramid fusion module. After convolution, the deepest layer of features is generated to predict the change mask.

[0078] The final prediction result is obtained by combining the five prediction results. The five prediction change masks obtained above are then concatenated and convolved to obtain the final prediction change mask.

[0079] M=Conv1([M1,M2,M3,M4,M5],1) (13)

[0080] Where M1, M2, M3, M4, and M5 represent the five prediction results obtained above, and M represents the final prediction result, i.e., the desired range of change.

[0081] IV. Calculating the Loss Function

[0082] The loss of the network model proposed in this embodiment of the invention consists of the loss of the attention pyramid fusion and the change mask map prediction network.

[0083] Calculate M respectively i , The three elements M (i = 1, 2, 3, 4, 5) and the actual region of change M GT The cross-entropy loss between the predictions is then used to compare each prediction with the actual change region M. GT The sum of the losses is used as the loss function of this network. That is, the loss function of the change detection method based on high-quality feature selection is:

[0084]

[0085] Among them, L Total This represents the total loss of the entire network. The three terms from left to right are: the first term represents the difference between the final predicted result and the actual change region M. GT The loss, the second term represents the difference between the initial prediction result and the actual change area M at each layer. GTThe loss, the third term represents the additional prediction results generated after attention pyramid fusion processing compared to the actual change region M. GT The loss.

[0086] In practical applications, pixel-level change detection is viewed as performing a binary classification on each pixel value. The prediction for each pixel has only two possibilities: changed or unchanged (1 or 0). Using the cross-entropy between the predicted result and the true label effectively determines the difference between them. The closer the two are, the smaller the cross-entropy; the greater the difference, the larger the cross-entropy. Introducing cross-entropy loss allows the model's predictions to gradually approach the true label during training.

[0087] V. Detection of changing areas

[0088] The loss function obtained in Part 4 is used to detect changing image regions, which solves the problem of edge defects or lack of prominence of changing regions caused by relying on a single feature, and improves the detection accuracy of changing image regions.

[0089] Example 2

[0090] The following is combined Figures 1-3 The feasibility of the scheme in Example 1 is verified, as detailed below:

[0091] Based on the characteristic that changing images contain rich feature information, the embodiments of the present invention are designed... Figure 1 The network structure shown can effectively determine the importance of various change-related features and combine them, overcoming the impact of incomplete foreground feature information or lack of prominence of change areas.

[0092] according to Figure 2 As shown in the network structure, the paired features extracted by the twin U-shaped network are used as input. The spliced ​​features and absolute difference features are obtained by using splicing, absolute difference and convolution operations. A feature pool of change-related features composed of three different types is constructed. Then, a weight matrix that can be continuously updated with network training is initialized to select and weight the relevant features. This can effectively combine multiple features to obtain high-quality change features.

[0093] Depend on Figure 3 As shown, the method proposed in this embodiment of the invention is called HQFSnet. Compared with the other seven existing change detection methods, the change detection result P1 obtained in this embodiment of the invention matches the real change region gt better, and achieves excellent results on different datasets.

[0094] Example 3

[0095] Based on the same inventive concept, embodiments of the present invention also provide a change detection device for high-quality feature selection, see [link to relevant documentation]. Figure 5 The device includes a processor 1 and a memory 2, wherein the memory 2 stores program instructions, and the processor 1 invokes the program instructions stored in the memory 2 to cause the device to perform the method steps of any of Embodiment 1:

[0096] The input image pairs are feature extracted using a Siamese U-shaped network to obtain image features at different scales, and these image features are used as input features to construct a feature pool. The Siamese U-shaped network has full attention enhancement characteristics and is based on the VGG16bn network.

[0097] Feature selection is performed based on the contribution level of different change-related features in the feature pool to obtain high-quality change features; pyramid-style feature fusion is performed on change features at different scales by combining attention.

[0098] Convolution operations are added to each layer of features output by pyramid feature fusion, and a change mask is generated for each layer. The five predicted change masks are concatenated and convolved to obtain the final predicted change mask.

[0099] A loss function is constructed based on attention pyramid fusion and the final predicted change mask, and the changed image regions are detected through the loss function.

[0100] Specifically, constructing a feature pool using image features as input features involves:

[0101] The paired features obtained from the Siamese network are used as input, and the concatenated features are obtained through concatenation and convolution operations. The absolute difference features are obtained through feature subtraction and convolution operations. The extracted paired features, concatenated features, and absolute difference features are used to construct a feature pool.

[0102] Furthermore, feature selection based on the contribution of input features to obtain high-quality variation features specifically involves treating these three different types of features as expert features, randomly initializing an updatable weight matrix, selecting features from the feature pool, and summing the weights based on their importance to obtain high-quality variation features.

[0103] Among them, the characteristics of high-quality changes are:

[0104]

[0105]

[0106] Among them, E k (,:,:,j) represents the feature of the j-th channel of the k-th expert, and W is a learnable parameter. This represents the weight value of the k-th expert feature in the j-th channel after normalization. Let represent the weighted value of the k-th expert feature in the j-th channel, and m represent the m-th feature among the k expert features. This represents the summation of the weighted transformation values ​​of the k expert features in the j-th channel.

[0107] Furthermore, the change mask is:

[0108]

[0109]

[0110] Among them, M i This represents the change mask for each layer of prediction. This indicates the use of features prior to prediction at each layer. Guided by features processed by the attention pyramid fusion module The features generated for this layer are used to predict the change mask for this layer. Conv1(·,1) indicates that the layer is processed by a convolution with a 1×1 kernel. This represents the features after processing by the attention pyramid fusion module. After convolution, the deepest layer of features is generated to predict the change mask.

[0111] It should be noted that the device descriptions in the above embodiments correspond to the method descriptions in the embodiments, and the embodiments of the present invention will not be repeated here.

[0112] The execution entities of the processor 1 and memory 2 mentioned above can be devices with computing functions such as computers, microcontrollers, and microcontrollers. In specific implementation, the embodiments of the present invention do not limit the execution entity and can select it according to the needs of actual application.

[0113] Data signals are transmitted between memory 2 and processor 1 via bus 3, which will not be described in detail in this embodiment of the invention.

[0114] Example 4

[0115] Based on the same inventive concept, embodiments of the present invention also provide a computer-readable storage medium, the storage medium including a stored program, which, when the program is running, controls the device where the storage medium is located to execute the method steps in the above embodiments.

[0116] The computer-readable storage medium includes, but is not limited to, flash memory, hard disk, solid-state drive, etc.

[0117] It should be noted that the description of the readable storage medium in the above embodiments corresponds to the description of the method in the embodiments, and the embodiments of the present invention will not be repeated here.

[0118] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of the present invention is generated.

[0119] A computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in or transmitted through a computer-readable storage medium. A computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available medium can be magnetic or semiconductor, etc.

[0120] Unless otherwise specified, the model numbers of the various devices in this embodiment of the invention are not limited, and any device that can perform the above functions is acceptable.

[0121] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0122] The above description is only a preferred 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 within the protection scope of the present invention.

Claims

1. A change detection method based on high-quality feature selection, characterized in that, The method includes: The input image pairs are feature extracted using a Siamese U-shaped network to obtain image features at different scales, and these image features are used as input features to construct a feature pool. The Siamese U-shaped network has full attention enhancement characteristics and is based on the VGG16bn network. Feature selection is performed based on the contribution level of different change-related features in the feature pool to obtain high-quality change features; pyramid-style feature fusion is performed on change features at different scales by combining attention. Convolution operations are added to each layer of features output by pyramid feature fusion, and a change mask is generated for each layer. The five predicted change masks are concatenated and convolved to obtain the final predicted change mask. A loss function is constructed based on attention pyramid fusion and the final predicted change mask. This loss function is then used to analyze the changes... Image regions are detected; The specific steps for constructing a feature pool using image features as input features are as follows: The paired features obtained from the Siamese network are used as input, and the concatenated features are obtained through concatenation and convolution operations. The absolute difference features are obtained through feature subtraction and convolution operations. The extracted paired features, concatenated features, and absolute difference features are used to construct a feature pool. Paired features, concatenated features, and absolute difference features are treated as expert features. An updatable weight matrix is ​​randomly initialized, and features in the feature pool are selected. The weights are summed according to their importance to obtain high-quality variation features. The high-quality change characteristics are: in, Let W represent the feature of the j-th channel of the k-th expert, where W is a learnable parameter. This represents the weight value of the k-th expert feature in the j-th channel after normalization. Let represent the weighted value of the k-th expert feature in the j-th channel, and m represent the m-th feature among the k expert features. This represents the summation of the weighted transformation values ​​of the k expert features in the j-th channel.

2. The change detection method based on high-quality feature selection according to claim 1, characterized in that, The change mask is: in, This represents the change mask for each layer of prediction. This indicates the use of features prior to prediction at each layer. Guided by features processed by the attention pyramid fusion module The features generated in this layer are used to predict the change mask of this layer. This indicates that processing is performed using a convolution with a 1×1 kernel. This represents the features after processing by the attention pyramid fusion module. After convolution, the deepest layer of features is generated to predict the change mask; This indicates that a feature upsampling operation is being performed.

3. A change detection device based on high-quality feature selection, characterized in that, The device includes a processor and a memory, the memory storing program instructions, the processor calling the program instructions stored in the memory to cause the device to perform the method according to any one of claims 1-2.

4. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, the computer program including program instructions that, when executed by a processor, cause the processor to perform the method described in any one of claims 1-2.

Images