Remote sensing image change detection method and system based on bidirectional enhanced Transform and channel reweighting

Abstract

The invention discloses a remote sensing image change detection method and system based on bidirectional enhanced Transform and channel reweighting in the technical field of image processing. In the coding part, a bidirectional enhanced self-attention (BEA) operation is provided, coding is carried out in the horizontal direction and the vertical direction, deep convolution is introduced to reinforce the relation of local contexts, and the perception ability of the network to global and local information is enhanced while the calculation complexity is reduced. In the feature fusion part, dual-time feature fusion is carried out by adopting a channel reweighting method, and the problem of insufficient or inconsistent feature information is solved by recalibrating each channel, suppressing noise features and enhancing semantic related features. According to the method, the performance of the network can be improved, and the change detection precision is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and in particular relates to a remote sensing image change detection method and system based on bidirectional enhanced Transformer and channel reweighting. Background Art

[0002] Change detection in remote sensing images is to quantitatively analyze and determine the characteristics and processes of surface changes from remote sensing data of different periods. Due to different imaging conditions and the complex and diverse characteristics of objects in the scene, high-resolution remote sensing images may show different spectral characteristics of objects with the same semantic concept at different time and spatial positions, which still poses great challenges to the change detection task. In order to accurately identify the changed areas in complex scenes, the change detection model needs to identify the high-level semantic information of the changed areas in the scene and distinguish the real changes from the complex unrelated changed areas.

[0003] By using self-attention to model the semantic relationship between pixel pairs in space and time, better detection results can be achieved. In 2020, Hao Chen et al. from Beihang University proposed STANet, a new twin network based on metric and spatiotemporal attention. It divides the network into different scales and embeds self-attention in each small area to model long-distance dependencies between pixels. It uses batch balanced contrast loss function to alleviate the problem of sample imbalance and help the network extract features rich in contextual information. However, the computational efficiency of the model is very low, the computational complexity is very high, and it grows quadratically with the number of pixels.

[0004] At present, many change detection methods based on deep convolutional neural networks (CNN) cannot effectively capture the long-distance dependencies between pixels. In this context, Hao Chen et al. from the Beijing University of Aeronautics and Astronautics proposed a bitemporal image transformer (BIT) detection algorithm, which achieved good results in change detection tasks. However, since each head of the multi-head attention in the transformer is only responsible for a subset of the input token sequence, when the channel dimension of the subset is small, the dot product of the query and key cannot form an information function, which will lead to reduced network performance. At the same time, BIT uses a direct subtraction feature fusion method that not only easily destroys the feature structure, but also may cause noise interference. Summary of the invention

[0005] The purpose of the present invention is to overcome the deficiencies in the prior art and provide a remote sensing image change detection method and system based on bidirectional enhanced Transformer and channel reweighting, so that the final change detection result of dual-phase multispectral remote sensing image is more accurate.

[0006] To achieve the above object, the technical solution adopted by the present invention is: a remote sensing image change detection method and system based on bidirectional enhanced Transformer and channel reweighting, comprising:

[0007] a. After the dual-time image is input, it first passes through the encoder. The encoder consists of a series of attention layers, each of which consists of a feedforward network and a bidirectional enhanced self-attention. The weights of the encoder are shared and used to calculate the multi-scale features of the dual-time image before and after the change. For each scale i, we can get the features of two streams: and

[0008] b. The characteristics of these two streams and Input into the channel weighted fusion module to encode the semantic change information and obtain the features at different scales i

[0009] c. These multi-scale features After being input into the decoder, the spatial resolution of the image is improved after a series of convolutional layers and transposed convolutional layers for feature fusion. These upsampled features are then passed to a prediction mask layer to obtain the semantic binary change map M of the dual-time image.

[0010] d. Throughout the training process, we used a hybrid loss function (Dice loss function and cross entropy loss function) to weaken the impact of unbalanced samples. By gradient updating the loss function, we can obtain the optimal change detection model after multiple rounds of training.

[0011] The bidirectional enhanced self-attention in step a is an improvement on the standard self-attention. On the one hand, it is used for global semantic extraction by squeezing operations in the horizontal and vertical directions. On the other hand, the model's perception of local spatial details is enhanced by the convolution-based detail enhancement kernel. The specific operations are as follows:

[0012] Pick is the encoder feature input to the i-th stage of BEA. In order to reduce the computational complexity while efficiently maintaining the global receptive field, we first perform a sparsity factor of α on the input feature F i Perform data-dependent sparse sampling operations to obtain α 2 Feature subset Here, the sparse factor α is used to control the complexity of sampling. The data-dependent sparse spatial subsampling operation of the feature is performed as follows:

[0013]

[0014] Get sparse sampling features After that, we perform a linear transformation as in equation (1) to obtain Q, K, and V. At this point, represents the input x, Next, we take the average of the feature maps Q, K, and V in the horizontal direction to achieve horizontal squeezing and get Q h , K h , Similarly, vertical extrusion is performed in the vertical direction to obtain Q v , K v , The relevant formula is as follows:

[0015]

[0016] Among them, x →(.) Indicates the permutation of the dimension of the tensor x, i W ∈R W ,i H ∈R H represents a vector whose elements are all 1. Through this operation, we can squeeze the global information onto one axis, thus easing the global semantic extraction as shown in the following equation:

[0017]

[0018] At this time, the time complexity required to compress Q, K, V is The time complexity required for the attention operation is Therefore, compared with the standard global self-attention operation, this operation can reduce the time complexity to O(H i W i ).

[0019] In addition, to help the model learn information at different positions in the input sequence, we also introduce position encoding. Come to Q h and K h These positional encodings are derived from learnable parameters In the vertical extrusion process, we use the same method to interpolate Application to Q v and K vThe attention calculation at this time is as follows:

[0020]

[0021] On the other hand, the present invention also adopts a series of operations to enhance the model's ability to perceive local details. First, Q, K and V are connected, and a 3×3 deep convolution kernel is passed to perceive local details and batch normalization is performed. Then, through a linear transformation with an activation function and batch normalization, Dimensional features are compressed to C i , and generate enhanced features. Finally, the enhanced features are fused with the features obtained through the attention mechanism to obtain a feature map with full resolution.

[0022] The channel weighted fusion module in step b adopts a new feature fusion strategy. This method dynamically weights the channels at each stage and adaptively adjusts the weights in the feature fusion process, thereby effectively suppressing noise interference and enhancing feature channels with significant semantic changes.

[0023] Specifically, we concatenate the pre-changed features and post-change characteristics Combine them and transform their shapes to get features Then the features After summing in the channel dimension, global average pooling is performed to obtain the feature vector v i , the steps are as follows:

[0024]

[0025] Among them, FC&ST represents feature concatenation and shape conversion, and CWS represents channel dimension summation. i After that, we transform the feature vector v i The input is fed into a shared Conv-ReLU layer to reduce the number of channels, and then these reduced features are fed into a separate 1×1 convolutional layer to obtain two streams. and The channel weights are as follows:

[0026]

[0027] Among them, ω1, ω2 and ω3 represent convolution operations, and They represent the non-normalized channel reweighting factors of the pre-change features and post-change features in stage i, respectively. Afterwards, these weights are normalized using the Softmax function as follows:

[0028]

[0029] Where j represents the channel index. After the channels are re-weighted, we perform feature fusion through multiplication and addition operations, and use gating operations to generate enhanced fused features As shown below:

[0030]

[0031] in, express The first part of , that is, the first tensor in the list, has the highest feature resolution and can retain the information in the input features to the greatest extent. δ represents a learnable gating coefficient ranging from 0 to 1, which is used to control the optimal ratio of the original fusion feature to the reweighted fusion feature.

[0032] In the decoding operation of step c, specifically, we use an MLP layer to decode the features of each scale The channel dimensions are unified and then upsampled to obtain C ebd represents the embedding dimension. Then these features Pass it into the MLP layer for splicing and fusion, and get Then upsample it to get Finally, we use another MLP layer to process the upsampled fused features. To predict semantic binary change masks for bi-temporal images Among them, N cl The value of is 2, indicating two categories: change and no change.

[0033] The mixed loss function (Dice loss function and cross entropy loss function) used in step d is as follows:

[0034] L(pred,gt)=L dice (pred,gt)+L wce (pred,gt) (15)

[0035] Among them, pred represents the predicted value and gt represents the true value.

[0036] A remote sensing image change detection method and system based on bidirectional enhanced Transformer and channel reweighting, comprising a processor and a storage device, wherein the storage device stores a plurality of instructions for the processor to load and execute the steps of any one of the above methods.

[0037] Compared with the prior art, the beneficial effects achieved by the present invention are as follows: In the remote sensing image change detection method and system based on bidirectional enhanced Transformer and channel reweighting described in this method, in the encoding part, we propose a bidirectional enhanced self-attention operation, which strengthens the connection of local context by encoding in the horizontal and vertical directions and introducing deep convolution, thereby reducing the computational complexity while enhancing the network's perception of global and local information. In the feature fusion part, we use the channel reweighting method to perform dual-time feature fusion, and by recalibrating each channel, suppressing noise features and enhancing semantically related features to solve the problem of insufficient or inconsistent feature information. This method makes the final change detection results of dual-phase multispectral remote sensing images more reliable and robust, and the detection accuracy is higher. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 is a schematic diagram of an implementation flow of a remote sensing image change detection method based on a bidirectional enhanced Transformer and channel reweighting according to an embodiment of the present invention, wherein Figure 1 (a), (b), and (c) represent the overall network process, the sub-coding block of the network, and the decoder of the network, respectively;

[0039] Figure 2 It is a bidirectional enhanced attention (Bidirectional-Enhanced Attention, BEA) of a remote sensing image change detection method based on a bidirectional enhanced Transformer and channel reweighting in an embodiment of the present invention;

[0040] Figure 3 It is a channel weighted fusion module (Channel Weighted Fusion Module, CWFM) of a remote sensing image change detection method based on a bidirectional enhanced Transformer and channel reweighting in an embodiment of the present invention;

[0041] Figure 4 is the LEVER optical remote sensing image change data set and detection results used in the embodiment of the present invention, wherein Figure 4 (a), (b), (c), (d), and (e) respectively represent the image before the change, the image after the change, the change reference image, the STANet detection result, the BIT detection result, and the detection result of the present invention. DETAILED DESCRIPTION

[0042] A remote sensing image change detection method based on bidirectional enhanced Transformer and channel reweighting includes: an overall framework, bidirectional enhanced attention and channel weighted fusion modules. In the encoding stage, we adopt an encoding structure that connects Transformer in multiple levels and propose a bidirectional enhanced self-attention. By encoding in the horizontal and vertical directions and introducing deep convolution to strengthen the connection of local context, the network's perception of global and local information is enhanced while reducing computational complexity. In the feature fusion part, we propose a channel weighted fusion module to fuse dual-time features. By recalibrating each channel, suppressing noise features and enhancing semantically related features, the problem of insufficient or inconsistent feature information can be solved. Finally, through a simple decoding process, the final semantic binary change map can be obtained.

[0043] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.

[0044] like Figure 1 (a) is a schematic diagram of the implementation process of remote sensing image change detection based on bidirectional enhanced Transformer and channel reweighting according to the present invention, which includes the following steps:

[0045] Step 1: After the dual-time image is input, it first passes through the encoder. The encoder consists of a series of sub-coding blocks. Figure 1 As shown in (b), the sub-encoding block we proposed is mainly composed of a bidirectional enhanced self-attention operation (BEA), a feed-forward network (FFN) and a normalization layer (Norm). is the encoder feature of the i-th stage input to the sub-encoding block. For each scale i, we can get the features of two streams: and

[0046] Among them, bidirectional enhanced self-attention (BEA) is the core of this step, and the specific structure is as follows Figure 2 It is an improvement over the standard self-attention. The standard global self-attention is shown below:

[0047] Q=W q x,K=W k x,V=W v x (6)

[0048]

[0049] Where x∈R H×W×C Indicates input, represents the learnable weights, and a(o) represents all positions on the feature map of o = (m, n). H×W×C When performing a standard self-attention operation, the time complexity may be O(H 2 W 2 (C v +C qk )), the memory required grows quadratically with the increase in the spatial resolution of the input image.

[0050] Therefore, in order to improve efficiency, the present invention proposes a bidirectional enhanced self-attention operation. On the one hand, squeezing operations are performed in the horizontal and vertical directions for global semantic extraction. On the other hand, the model's perception of spatial local details is enhanced by a convolution-based detail enhancement kernel. The specific operation is as follows:

[0051] Pick is the encoder feature input to the i-th stage of BEA. In order to reduce the computational complexity while efficiently maintaining the global receptive field, we first perform a sparsity factor of α on the input feature F i Perform data-dependent sparse sampling operations to obtain α 2 Feature subset Here, the sparse factor α is used to control the complexity of sampling. The data-dependent sparse spatial subsampling operation of the feature is performed as follows:

[0052]

[0053] Get sparse sampling features After that, we perform a linear transformation as in equation (1) to obtain Q, K, and V. At this point, represents the input x, Next, we take the average of the feature maps Q, K, and V in the horizontal direction to achieve horizontal squeezing and get Q h , K h , Similarly, vertical extrusion is performed in the vertical direction to obtain The relevant formula is as follows:

[0054]

[0055] Among them, x →(.) Indicates the permutation of the dimension of the tensor x, i W ∈R W ,i H ∈R H represents a vector whose elements are all 1. Through this operation, we can squeeze the global information onto one axis, thus easing the global semantic extraction as shown in the following equation:

[0056]

[0057] At this time, the time complexity required to compress Q, K, V is The time complexity required for the attention operation is Therefore, compared with the standard global self-attention operation, this operation can reduce the time complexity to O(H i W i ).

[0058] In addition, to help the model learn information at different positions in the input sequence, we also introduce position encoding. Come to Q h and K h These positional encodings are derived from learnable parameters In the vertical extrusion process, we use the same method to interpolate Application to Q v and K v The attention calculation at this time is as follows:

[0059]

[0060] like Figure 2 As shown in the following path, the present invention also takes a series of operations to enhance the model's ability to perceive local details. First, Q, K, and V are connected, a 3×3 deep convolution kernel is passed in to perceive local details, and batch normalization is performed. Then, through a linear transformation with an activation function and batch normalization, dimensional features are compressed to C i , and generate enhanced features. Finally, the enhanced features are fused with the features obtained through the attention mechanism to obtain a feature map with full resolution.

[0061] In addition, during the training phase of the BEA operation, we also used the Dropout regularization technique to reduce the overfitting of the model. After completing the BEA operation, we can obtain the feature F of each stage through the following operations: i :

[0062]

[0063] Here, F i Different flows are divided into pre-change features and post-change characteristics

[0064] Step 2: Combine the features of these two streams and Input into the channel weighted fusion module (CWFM), encode the semantic change information, and obtain the features at different scales i

[0065] Among them, the structure of the channel weighted fusion module (CWFM) proposed in the present invention is as follows: Figure 3 As shown in the figure, a CWFM module is introduced in each stage of the encoder to achieve feature fusion of the four stages. The specific operations are as follows:

[0066] First, we concatenate the pre-changed features and post-change characteristics Combine them and transform their shapes to get features Then the features After summing in the channel dimension, global average pooling is performed to obtain the feature vector v i , the steps are as follows:

[0067]

[0068] Among them, FC&ST represents feature concatenation and shape conversion, and CWS represents channel dimension summation. i After that, we transform the feature vector v i Input to the shared Cony-ReLU layer to reduce the number of channels, and then pass these reduced features to a separate 1×1 convolutional layer to obtain two streams and The channel weights are as follows:

[0069]

[0070] Among them, ω1, ω2 and ω3 represent convolution operations, and They represent the non-normalized channel reweighting factors of the pre-change features and post-change features in stage i, respectively. Afterwards, these weights are normalized using the Softmax function as follows:

[0071]

[0072] Where j represents the channel index. After the channels are re-weighted, we perform feature fusion through multiplication and addition operations, and use gating operations to generate enhanced fused features As shown below:

[0073]

[0074] in, express The first part of , that is, the first tensor in the list, has the highest feature resolution and can retain the information in the input features to the greatest extent. δ represents a learnable gating coefficient ranging from 0 to 1, which is used to control the optimal ratio of the original fusion feature to the reweighted fusion feature.

[0075] Step 3: Combine these multi-scale features After being input into the decoder, the spatial resolution of the image is improved after a series of convolutional layers and transposed convolutional layers for feature fusion. After that, these upsampled features are passed to a prediction mask layer, and finally the semantic binary change map M of the dual-time image is obtained. The structure of the decoder is as follows Figure 1 (c) as shown.

[0076] Specifically, we use an MLP layer to transform the features of each scale The channel dimensions are unified and then upsampled to obtain C ebd represents the embedding dimension. Then these features Pass it into the MLP layer for splicing and fusion, and get Then upsample it to get Finally, we use another MLP layer to process the upsampled fused features. To predict semantic binary change masks for bi-temporal images Among them, N cl The value of is 2, indicating two categories: change and no change.

[0077] Step 4: During the entire training process, we used a mixed loss function (Dice loss function and cross entropy loss function) to weaken the impact of unbalanced samples. The specific formula is as follows:

[0078] L(pred,gt)=L dice (pred,gt)+L wce (pred,gt) (15)

[0079] Among them, pred represents the predicted value and gt represents the true value.

[0080] The Dice loss function is calculated as follows:

[0081]

[0082] Among them, t n represents the true label category of pixel n. When n is a pixel of the same class, t n The value is 0. When n is a change pixel, t n The value is 1; nrepresents the predicted probability that pixel n belongs to the change class.

[0083] The cross entropy loss function is calculated as follows:

[0084]

[0085] Among them, t m represents the true label category of pixel m, y m represents the predicted probability that pixel m belongs to the change class.

[0086] The present invention is further described in detail below in conjunction with specific experimental data. The experimental data used in this experiment is LEVER change detection data, and the image size is cropped to 256×256. The experimental platform is Win11 system, the graphics card is NVIDIAGeForce RTX 3060, the deep learning framework used is PyTorch framework, and the cuDNN library and CUDA computing architecture are used for training acceleration. During the training process, data enhancement techniques are applied, including Gaussian blur, random scaling, random flipping, random cropping, and random color transformation. In addition, this paper uses a mixed loss function (Dice loss function and cross entropy loss function) and AdamW optimizer for model training, with a weight decay equal to 0.01 and a β value equal to (0.9, 0.999). In the experiment, the batch size is set to 4, the initial learning rate is set to 0.0004, the epoch is set to 200, the sparse factor α in the sparse sampling operation is set to 4, and a linear decay operation is used to reduce the learning rate until it drops to 0 in the last period.

[0087] In order to verify the effectiveness of the present invention, the change detection method of the present invention is compared with the following change detection methods:

[0088] (1) STANet detection method proposed by Hao Chen et al. [“A Spatial-Temporal Attention-Based Method and a New Dataset for Remote Sensing Image Change Detection.” (Remote Sensing, 2020, 12(10)]

[0089] (2) BIT detection method (GAN) proposed by Hao Chen et al. [Hao Chen et al. Detection method proposed in the article "Remotesensing image change detection with transformers. (IEEE Transactions on Geoscience and Remote Sensing, 2022, 60]

[0090] (3) The method of the present invention.

[0091] The detection performance is measured using precision, recall, and F1 score indicators. F1 is a comprehensive evaluation indicator for change detection. The closer its value is to 1, the better the performance of the change detection method. The detection results are shown in Table 1.

[0092] Table 1. Change detection results performance on LEVIR dataset (%)

[0093]

[0094] As can be seen from Table 1, the F1 score of the detection method proposed in the present invention is the largest, which is closer to 1 than the other three detection algorithms. In addition, the Precision and Recall of the present invention are the largest among the comparison algorithms and are closer to 1. In summary, the performance of the change detection algorithm of the present invention is better than the other three detection methods, which shows that the change detection method proposed in the present invention is effective.

[0095] Figure 4 (a) and (b) are the before and after phase images in the LEVER dataset. Figure 4 (c) is the reference image for change detection. Figure 4 (d), (e), and (f) are the change detection results of the STANet algorithm, the BIT algorithm, and the present invention, respectively. Figure 4 Compared with the reference image, the detection effect of the algorithm proposed in the present invention is the best in terms of visual effect.

[0096] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0097] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0098] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0099] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0100] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A remote sensing image change detection method based on bidirectional enhanced Transformer and channel reweighting, characterized in that: include: a. After the dual-time image is input, it first passes through the encoder. The encoder consists of a series of attention layers, each of which consists of a feedforward network and a bidirectional enhanced self-attention (BEA). The weights of the encoder are shared and used to calculate the multi-scale features of the dual-time image before and after the change. For each scale i, we can get the features of two streams: and b. The characteristics of these two streams and Input into the channel weighted fusion module (CWFM), encode the semantic change information, and obtain the features at different scales i c. These multi-scale features After being input into the decoder, the spatial resolution of the image is improved after a series of convolutional layers and transposed convolutional layers for feature fusion. After that, these upsampled features are passed to a prediction mask layer, and finally the semantic binary change map M of the dual-time image is obtained. d. Throughout the training process, we used a hybrid loss function (Dice loss function and cross entropy loss function) to weaken the impact of unbalanced samples. By gradient updating the loss function, we can obtain the optimal change detection model after multiple rounds of training.

2. According to claim 1, the bidirectional enhanced self-attention (BEA) in step a can be used for global semantic extraction by performing squeezing operations in the horizontal and vertical directions. The specific operations are as follows: Pick is the encoder feature input to the i-th stage of BEA. In order to reduce the computational complexity while efficiently maintaining the global receptive field, we first perform a sparsity factor of α on the input feature F i Perform data-dependent sparse sampling operations to obtain α 2 Feature subset Here, the sparse factor α is used to control the complexity of sampling. The data-dependent sparse spatial subsampling operation of the feature is performed as follows: Get sparse sampling features After that, we perform a linear transformation as in equation (1) to obtain Q, K, and V. At this point, represents the input x, Next, we take the average of the feature maps Q, K, and V in the horizontal direction to achieve horizontal squeezing and get Q h , K h , Similarly, vertical extrusion is performed in the vertical direction to obtain Q v , K v , The relevant formula is as follows: in, x →(.) Indicates the permutation of the dimension of the tensor x, i W ∈R W ,i H ∈R H represents a vector whose elements are all 1. Through this operation, we can squeeze the global information onto one axis, thus easing the global semantic extraction as shown in the following equation: At this time, the time complexity required to compress Q, K, V is The time complexity required for the attention operation is Therefore, compared with the standard global self-attention operation, this operation can reduce the time complexity to O(H i W′). In addition, to help the model learn information at different positions in the input sequence, we also introduce position encoding. Come to Q h and K h These positional encodings are derived from learnable parameters In the vertical extrusion process, we use the same method to interpolate Application to Q v and K v The attention calculation at this time is as follows: On the other hand, the present invention also adopts a series of operations to enhance the model's ability to perceive local details. First, Q, K and V are connected, a 3×3 deep convolution kernel is passed to perceive local details, and batch normalization is performed. Then, through a linear transformation with an activation function and batch normalization, dimensional features are compressed to C i , and generate enhanced features. Finally, the enhanced features are fused with the features obtained through the attention mechanism to obtain a feature map with full resolution.

3. According to claim 1, the channel weighted fusion module (CWFM) in step b can dynamically weight the channels of each stage and adaptively adjust the weights in the feature fusion process, thereby effectively suppressing noise interference and enhancing feature channels with significant semantic changes. Specifically, we concatenate the pre-changed features and post-change characteristics Combine them and transform their shapes to get features Then the features After summing in the channel dimension, global average pooling is performed to obtain the feature vector v i , the steps are as follows: in, FC&ST stands for feature concatenation and shape conversion, and CWS stands for channel dimension summation. i After that, we transform the feature vector v i The input is fed into a shared Conv-ReLU layer to reduce the number of channels, and then these reduced features are passed into a separate 1×1 convolutional layer to obtain two streams. and The channel weights are as follows: Among them, ω1, ω2 and ω3 represent convolution operations, and They represent the non-normalized channel reweighting factors of the pre-change features and post-change features in stage i, respectively. Afterwards, these weights are normalized using the Softmax function as follows: Where j represents the channel index. After the channels are re-weighted, we perform feature fusion through multiplication and addition operations, and use gating operations to generate enhanced fused features As shown below: in, express The first part of , that is, the first tensor in the list, has the highest feature resolution and can retain the information in the input features to the greatest extent. δ represents a learnable gating coefficient ranging from 0 to 1, which is used to control the optimal ratio of the original fusion feature to the reweighted fusion feature.

4. According to claim 1, the decoding operation in step c, specifically, we use an MLP layer to decode the features of each scale The channel dimensions are unified and then upsampled to obtain C ebd represents the embedding dimension. Then these features Pass it into the MLP layer for splicing and fusion, and get Then upsample it to get Finally, we use another MLP layer to process the upsampled fused features. To predict semantic binary change masks for bi-temporal images in, N cl The value of is 2, indicating two categories: change and no change.

5. According to claim 1, the hybrid loss function (Dice loss function and cross entropy loss function) used in step d is used for gradient updating, and the specific formula is as follows: L(pred,gt)=L dice (pred,gt)+L wce (pred,gt) (15) in, Pred represents the predicted value, and gt represents the true value.

6. A remote sensing image change detection system based on bidirectional enhanced Transformer and channel reweighting, characterized in that: It comprises a processor and a storage device, wherein the storage device stores a plurality of instructions for the processor to load and execute the steps of the method according to any one of claims 1 to 6.

Images