Infrared small target detection method based on multi-scale mutual attention mechanism

Abstract

The invention discloses an infrared small target detection method based on a multi-scale mutual attention mechanism, relates to the technical field of infrared small target detection, and solves the problem that in the prior art, refined capture of multi-shape and multi-scale information is difficult to realize. And feature enhancement lacks collaborative utilization of local and global contexts of space-channel dimensions, and complementarity and relevance among features of different scales cannot be fully mined. According to the method, a target detection network model based on multi-aspect feature extraction and a mutual attention mechanism is established, and a multi-shape receptive field extraction module, a space-channel-global-local joint re-extraction module and a multi-scale mutual attention mechanism module are designed in the network model; the shape features of the target are extracted and the target features are captured in a deeper level while the receptive field is expanded, and the multi-scale information of the target is fully fused by extracting more useful target feature information, so that the superior detection of the infrared small target is realized.

Description

Technical Field

[0001] This invention relates to the field of infrared small target detection technology, specifically to an infrared small target detection method based on a multi-scale mutual attention mechanism. Background Technology

[0002] Significant progress has been made in the field of infrared small target detection (IRSTD), but the task is still limited by three core challenges, which severely restrict detection accuracy and efficiency:

[0003] 1. Insufficient granularity of feature representation. Current IRSTD methods largely rely on stacked convolutional kernels to extract local features, and conventional convolutional operations struggle to adaptively capture fine-grained features such as the shape and scale of the target. This limitation in representation directly leads to low feature discrimination between the target and complex backgrounds, creating potential for false positives and false negatives in subsequent detection tasks.

[0004] 2. The effectiveness and specificity of feature enhancement are lacking. Existing research often uses linear squeeze attention mechanisms for feature enhancement, which, while strengthening target features, tends to amplify background interference information similar to the target features simultaneously. More importantly, it cannot effectively aggregate the global contextual information of small targets, resulting in insufficient semantic understanding of the target by the model, thus leading to false warnings.

[0005] 3. Insufficient depth in multi-scale feature fusion. Current construction of multi-scale feature layers relies heavily on traditional methods such as downsampling and wavelet transform, while feature fusion generally employs simple addition or direct splicing strategies. This fusion method struggles to retain the unique target discrimination information in features at different scales and fails to efficiently enhance the target features shared across scales, ultimately resulting in the loss of deep semantic features and making it difficult to adapt to the scale-variable characteristics of small infrared targets.

[0006] While existing infrared small target detection methods are mostly based on the UNet encoder-decoder architecture, they still face three core challenges in complex scenarios: First, small targets have sparse feature dimensions, making it difficult for existing feature extraction modules to achieve refined capture of multi-shape and multi-scale information; second, infrared images have low signal-to-noise ratios and strong background interference, and feature enhancement lacks the coordinated utilization of local and global contexts in the spatial-channel dimension; and third, multi-scale feature fusion is mostly a simple stitching or weighted summation, failing to fully explore the complementarity and correlation between features of different scales. Summary of the Invention

[0007] To address the challenges of existing technologies in achieving refined capture of multi-shape and multi-scale information, the lack of collaborative utilization of local and global context in spatial-channel dimensions by feature enhancement, and the failure to fully exploit the complementarity and correlation between features at different scales, this invention provides an infrared small target detection method based on a multi-scale mutual attention mechanism.

[0008] An infrared small target detection method based on a multi-scale mutual attention mechanism is proposed, which is implemented by the following steps:

[0009] Step 1: Obtain the infrared small target detection dataset;

[0010] Step 2: Perform data augmentation on the dataset detected in Step 1 to obtain the processed dataset, and divide the dataset into training set, validation set and test set;

[0011] Step 3: Establish a target detection network model based on multi-faceted feature extraction and mutual attention mechanism; the network model includes a multi-shape receptive field extraction module for feature extraction, a spatial-channel-global-local joint re-extraction module for feature enhancement and re-extraction, and a multi-scale mutual attention mechanism module for feature fusion; input the training set into the network model for training;

[0012] Step 4: Load the validation set into the network model, calculate the loss function, backpropagate, update the learning rate, and save the network model parameters;

[0013] Step 5: Load the validation set and the network model parameters saved in Step 4 into the network model, obtain the detection results, and validate the network model; after multiple rounds of training, obtain the optimal model parameters.

[0014] Step 6: Load the test set and optimal model parameters into the network model to obtain the results of infrared small target detection.

[0015] The beneficial effects of this invention are:

[0016] 1. The detection method described in this invention proposes an infrared small target detection algorithm that combines multi-faceted feature extraction and a multi-scale cross-attention mechanism. By extracting more useful target feature information and fully integrating the multi-scale information of the target, superior detection of infrared small targets is achieved.

[0017] 2. The detection method of the present invention designs a multi-shape receptive field extraction module, which uses deformable dilated convolution and convolution fusion structure to extract the shape features of the target while expanding the receptive field.

[0018] 3. The detection method described in this invention incorporates a spatial-channel-global-local joint re-extraction module to analyze the local, global, and dimensional information of the target. It employs a multi-directional feature extraction structure, including local-spatial-global channels, global-spatial-local channels, and local-spatial-global channel structures, to capture target features at a deeper level.

[0019] 4. The detection method described in this invention incorporates a multi-scale cross-attention mechanism module, which integrates feature information from different scales to maximize feature utilization. Attached Figure Description

[0020] Figure 1 This is a flowchart of an infrared small target detection method based on a multi-scale mutual attention mechanism as described in this invention;

[0021] Figure 2 This is a schematic diagram of a target detection network model based on multi-feature extraction and mutual attention mechanism. Figure 3 This is a schematic diagram of the multi-shape receptive field extraction module.

[0022] Figure 4 This is a schematic diagram of the spatial-channel-global-local joint extraction module.

[0023] Figure 5 This is a schematic diagram of the local-spatial-local-channel module.

[0024] Figure 6 This is a schematic diagram of the infrared small target detection method based on a multi-scale mutual attention mechanism according to the present invention.

[0025] Figure 7 These are diagrams illustrating the effects of the detection method of this invention and existing technologies. Detailed Implementation

[0026] Combination Figures 1 to 7 This embodiment describes an infrared small target detection method based on a multi-scale mutual attention mechanism. This method features an innovative design on the UNet architecture, constructing a three-stage optimization framework of refined extraction, collaborative enhancement, and deep fusion to achieve small target detection. The specific detection method is as follows:

[0027] Step 1: Obtain the infrared small target detection dataset;

[0028] Step 2: Adjust image size and perform data augmentation on the dataset, including randomly cropping the images to 256×256 pixels, randomly flipping the images horizontally and vertically by 50%, and randomly transposing the images by 50%. Obtain the processed dataset, and divide the dataset into 80% training set, 10% validation set, and 10% test set.

[0029] Step 3: Establish a target detection network model based on multi-faceted feature extraction and mutual attention mechanism;

[0030] like Figure 2As shown, the network model includes a multi-shape receptive field extraction module for feature extraction, a spatial-channel-global-local joint re-extraction module for feature enhancement and re-extraction, and a multi-scale mutual attention mechanism module for feature fusion.

[0031] The multi-shape receptive field extraction module serves as the backbone of the network's feature extraction. This module expands the receptive field by dividing the feature map into four branches and fusing each branch sequentially, thereby improving the extraction of target size information. Within each branch, deformable dilated convolution (DCNConv) is used to process each feature layer, as DCNConv can capture the target's shape information. Therefore, the multi-shape receptive field extraction module can better extract size and shape cues from the target.

[0032] like Figure 3 As shown, in this embodiment, the input image enters the multi-shape receptive field extraction module for feature extraction;

[0033] First, the image is subjected to 3×3 convolution, batch normalization, and ReLU activation function operation (Conv / k3+BN+ReLU) to obtain the feature map;

[0034] Then, the feature map is split into four blocks for differential feature extraction. The second block is processed by a deformable dilated convolution (DCNConv) with a kernel of 3 and a dilation factor of 2, followed by a normalization layer (BN) and an activation function (ReLU), and then adjusted with learnable parameters. The result is fused with the third block, processed by a deformable dilated convolution (DCNConv) with a kernel of 3 and a dilation factor of 2, followed by a normalization layer (BN), an activation function (ReLU), and then adjusted with learnable parameters. The result is fused with the fourth block, processed by a deformable dilated convolution (DCNConv) with a kernel of 3 and a dilation factor of 3, followed by a normalization layer (BN), an activation function (ReLU), and then adjusted with learnable parameters. Finally, the three processed blocks are concatenated with the first block, then fused with the feature map before splitting, and then processed by a 3×3... After convolution, batch normalization, and ReLU activation function operations (Conv / k3+BN+ReLU), channel attention and spatial attention mechanisms are used for enhancement to obtain the image with extracted features. The learnable parameters can obtain optimal modulation weights through model training, and by weighted multiplication and fusion of features from various branches, more comprehensive intermediate features are obtained.

[0035] The spatial-channel-global-local joint re-extraction module serves as a feature re-extraction component of the network model, re-extracting local and global information of the target across spatial and channel dimensions; this process obtains more comprehensive target information from multiple aspects.

[0036] like Figure 4 and Figure 5 As shown, in this embodiment, the spatial-channel-global-local joint re-extraction module includes: a local-spatial-local-channel module (performing feature extraction of image height and width respectively), a global channel extraction module (performing feature extraction of image height and width respectively), and a local channel extraction module;

[0037] Two image patches are processed separately through a local-spatial-local-channel module, a global channel extraction module, and a local channel extraction module, then dimensionally stitched together. This result is added to the image before splitting, yielding multiple local channel images processed by the spatial-channel-global-local joint extraction module. The specific implementation process is as follows:

[0038] The intermediate features extracted by the multi-shape receptive field extraction module are used as input to the spatial-channel-global-local joint re-extraction module. First, convolution, batch normalization, and ReLU activation function operations (Conv / k1+BN+ReLU) are performed. Then, the image is split into two image blocks. One image block enters the local-spatial-local-channel module (for image width feature extraction), as shown below. Figure 5 As shown, the image passes through a convolution with kernel 1, batch normalization, and ReLU activation (Conv / k1+BN+ReLU), then a convolution with kernel (1,3), batch normalization, and ReLU activation (Conv / k(1,3)+BN+ReLU), then a dilated convolution with kernel (1,3) and dilation factor 2, batch normalization, and ReLU activation (Conv / k(1,3)d2+BN+ReLU), and finally a convolution with kernel 1, batch normalization, and ReLU activation (Conv / k1+BN+ReLU). Finally, it passes through a sigmoid normalization function, multiplies with the image patch, and then adds the results to complete the residual connection.

[0039] Then, the output is input into the global channel extraction module (for image width feature extraction). First, the average and maximization operations of the channel dimensions are performed on it respectively. Then, the results of the average and maximization operations are concatenated, followed by convolution with kernel (1,3), batch normalization, and ReLU activation function operation (Conv / k(1,3)+BN+ReLU). Then, it is passed through a sigmoid normalization function, multiplied with the image block input to the global channel extraction module, and then added to complete the residual connection.

[0040] Finally, the image patch output by the global channel extraction module is input into the local channel extraction module. First, it is downsampled to 1×1 size in the spatial dimension, and then the overall dimension is adjusted from (B,C,1,1) to (B,C,1), where B represents the image batch and C represents the image channel dimension. Then, it undergoes a one-dimensional convolution with a kernel of 3, batch normalization, and ReLU activation function operation (Convd / k3BN+ReLU). The overall dimension is then adjusted from (B,C,1) to (B,C,1,1), and then passed through a sigmoid normalization function. Finally, it is multiplied and added with the image patch input to the local channel extraction module to complete the residual connection.

[0041] Another image patch enters the local-spatial-local-channel module (for image height feature extraction), undergoing convolution with kernel 1, batch normalization, and ReLU activation (Conv / k1+BN+ReLU), followed by convolution with kernel (3,1), batch normalization, and ReLU activation (Conv / k(3,1)+BN+ReLU), then dilated convolution with kernel (3,1) and dilation factor 2, batch normalization, and ReLU activation (Conv / k(3,1)d2+BN+ReLU), and finally convolution with kernel 1, batch normalization, and ReLU activation (Conv / k1+BN+ReLU). After passing through a sigmoid normalization function, the patch is multiplied and added together to complete the residual connection.

[0042] Then, the output is input into the global channel extraction module (for image height feature extraction). First, the average and maximization operations are performed on the channel dimensions respectively. Then, the results of the average and maximization operations are concatenated, followed by convolution with a kernel of (3,1), batch normalization, and ReLU activation function operation (Conv / k(3,1)+BN+ReLU). After passing through a sigmoid normalization function, it is multiplied with the image block input to the global channel extraction module and then added to complete the residual connection.

[0043] Finally, the image patch output by the global channel extraction module is input into the local channel extraction module. First, it is downsampled to 1×1 size in the spatial dimension, and then the overall dimension is adjusted from (B,C,1,1) to (B,C,1), where B represents the image batch and C represents the image channel dimension. Then, it undergoes a one-dimensional convolution with a kernel of 3, batch normalization, and ReLU activation function operation (Convd / k3BN+ReLU). The overall dimension is then adjusted from (B,C,1) to (B,C,1,1), and then passed through a sigmoid normalization function. Finally, it is multiplied and added with the image patch input to the local channel extraction module to complete the residual connection.

[0044] like Figure 6 As shown, the multiple local channel images are input to the multi-scale mutual attention mechanism module. This module downsamples the multi-scale features to the same size, rearranges them, and performs matrix multiplication to obtain a feature fusion matrix. The feature fusion matrix is ​​then multiplied with the multi-scale features to obtain the final multi-scale mutual attention mechanism. This is used to modulate the feature layer, followed by channel attention and spatial attention mechanisms, and then upsampled to obtain the multi-scale feature fusion information, which aids in subsequent feature recovery. The specific implementation process is as follows:

[0045] The multi-scale mutual attention mechanism module divides multiple local channel images into input image 1, input image 2, input image 3 and input image 4 from largest to smallest;

[0046] First, input images 1 through 4 are uniformly downsampled and resized to 32×32. Then, the four resized input images are stitched together. Finally, the dimensions of the four resized input images are changed from (B, C) i Adjust ,32,32) to (B,C) i (,1024), where C i For the corresponding channel dimensions of the four input images, i can take the values ​​1, 2, 3, and 4. The dimensions of the concatenated image range from (B, C) a Adjust (32,32) to (B,1024,C) a ), where C a It is the sum of the channel dimensions of the four input images. Then, the four adjusted images are successively multiplied by the concatenated adjusted image to obtain four dimensions (B, C). i C a Multi-scale feature fusion factor;

[0047] Then, each multi-scale feature fusion factor is transposed, normalized, and processed using the softmax function (T / IN / Softmax) to obtain four dimensions (B, C). a C i The normalized multi-scale feature transpose fusion factor and normalization, softmax function processing (IN / Softmax) yield four dimensions (B, C). i C a The normalized multi-scale feature fusion factor is obtained by multiplying the normalized multi-scale feature transpose fusion factor with the four adjusted images to obtain four (B, C) normalized multi-scale feature transpose fusion factors. a The feature matrix (B, C) is then multiplied by the normalized multi-scale feature fusion factor to obtain four (B, C) features. i The fused matrix of (1024);

[0048] Then adjust the merged matrix to have dimensions (B, C).i The image is fused from the input image (32,32), and then channel attention and spatial attention modulation operations are performed on the fused image. Finally, an upsampling operation is performed to obtain four output fused images of the same size as the input image, thus completing the fusion of the multi-scale mutual attention mechanism module.

[0049] The image fused by the multi-scale mutual attention mechanism module is input into the existing image restoration module (CSFM) to restore the predicted target result.

[0050] like Figure 2 and Figure 6 As shown. In this embodiment, the input image undergoes five multi-shape receptive field extraction modules and four spatial-channel-global-local joint re-extraction modules to obtain multi-scale feature information. The multi-scale feature layer is then input into the multi-scale mutual attention mechanism module. For example... Figure 2 As shown, the overall network design consists of 5 multi-shape receptive field extraction modules, 4 spatial-channel-global-local joint re-extraction modules, 1 multi-scale mutual attention mechanism module, 1 upsampling module, and 1 prediction module. The input image first passes through the multi-shape receptive field extraction module to obtain input image 1; after passing through the multi-shape receptive field extraction module, it is enhanced by the spatial-channel-global-local joint re-extraction module; after passing through the multi-shape receptive field extraction module again, it obtains input image 2; after being enhanced by the spatial-channel-global-local joint re-extraction module again, it passes through the multi-shape receptive field extraction module to obtain input image 3; after being enhanced by the spatial-channel-global-local joint re-extraction module again, it passes through the multi-shape receptive field extraction module to obtain input image 4; input images 1-4 are then input into the multi-scale mutual attention mechanism module to obtain output images 1-4; the input images... 4. The image is then input into the spatial-channel-global-local joint extraction module for enhancement, and together with the output image 4, it is input into the existing image restoration module (CSFM) to obtain the restored image 3. The restored image 3 and the output image 3 are input into the existing image restoration module (CSFM) to obtain the restored image 2. The restored image 2 and the output image 2 are input into the existing image restoration module (CSFM) to obtain the restored image 1. The restored image 1 and the output image 1 are stitched together, and then passed through the upsampling module to obtain an output image of the same size as the input image. Finally, the prediction module is used to obtain a binary prediction result with the target being 1 and the background being 0.

[0051] The multi-scale mutual attention mechanism module described in this embodiment solves the problem of missing information at different scales during the fusion process by using methods such as feature fusion matrix transformation, compensates for the unique information in multi-scale features, and enhances the sharing of information.

[0052] Step 4: Set the initial learning rate of the network model to 0.0005, train for a total of 400 epochs, reduce the learning rate to 0.00005 at the 200th epoch, and reduce it to 0.000005 at the 300th epoch; establish the loss function SoftIoULoss by calculating the intersection-over-union ratio (IoU) of the predicted image and the real image; set the initial detection rate, normalized IoU, F1 score, and false alarm rate.

[0053] Step 5: Load the enhanced dataset from Step 1 into the object detection network model based on multi-feature extraction and mutual attention mechanism, calculate the loss function, backpropagate, update the learning rate, and save the model parameters.

[0054] Step 6: Load the validation set and saved model parameters from the dataset into the model, obtain the detection results, and validate the network model.

[0055] Step 7: After multiple rounds of training, obtain the optimal model parameters;

[0056] Step 8: Load the test set and optimal model parameters from the dataset into the model to obtain the detection results.

[0057] The detection method described in this embodiment addresses the problem of insufficient fine-grained feature capture by proposing a multi-shape receptive field extraction module, such as... Figure 3 As shown: By combining the Res2Net network structure with multi-rate dilated convolution, a dynamic receptive field adaptive adjustment mechanism is constructed to achieve accurate extraction of multi-shape contours and multi-scale size features of infrared small targets, providing a highly recognizable basic feature representation for subsequent feature processing. This module is embedded in the backbone layer and feature extraction stage of the UNet network.

[0058] The detection method described in this embodiment addresses the lack of effectiveness and specificity in feature enhancement by designing a spatial-channel-global-local joint re-extraction module. This module constructs a series structure of a local-spatial-local-channel module, a global channel extraction module, and a local channel extraction module. Through the interaction and enhancement of cross-dimensional contextual information, it effectively suppresses background noise interference, improves the recognizability of target features, and reduces the risk of false detections and misdetections. This module is applied in the feature enhancement and re-extraction stage.

[0059] The detection method described in this embodiment addresses the problem of insufficient depth in multi-scale feature fusion by designing a multi-scale mutual attention mechanism module. By introducing inter-scale mutual attention weight calculation, it adaptively mines the unique and common information of the target in features at different scales, realizing complementary fusion and enhancement of multi-scale features. The fused features are then passed to the decoder, providing more comprehensive feature support for subsequent feature recovery and target localization. This module is deployed in the feature fusion stage between the encoder and decoder.

[0060] The re-extraction module described in this embodiment optimizes the traditional UNet architecture from three dimensions: feature extraction accuracy, feature enhancement effectiveness, and feature fusion depth. It constructs a dedicated network framework adapted to the characteristics of infrared small target detection tasks, which is significantly different from the design ideas of existing methods in encoder and decoder.

[0061] like Figure 7 As shown in this embodiment, Figure 7 The first row contains four original images, each representing a different scenario: (a) an infrared image of a drone target against a background of distant city buildings; (b) an infrared image of a drone target against a background of nearby city buildings; (c) an infrared image of a drone target against a background of complex forest; and (d) an infrared image of an aircraft target against a background of the sky. The original images are then fed into the trained network model, which outputs the detection results, such as... Figure 7 The second row uses black to represent the background and white to represent the target area. The white area is magnified and displayed in the four corners of the image for easy observation of the detection effect. The red box highlights the target area. Figure 7 The third row shows the detection results of the existing advanced method (ALCNet). The blue box represents the case where no target was detected, and the green box represents the case where the background was detected as the target. In (a) and (b), the ALCNet method failed to detect the target. In (c), the ALCNet method detected the background as the target. In (d), the ALCNet method failed to detect the entire aircraft. In comparison, the method of the present invention has better detection performance. Figure 7 The fourth row shows the actual target mask image, with white representing the target area and black representing the background area. For easier observation, the white target area is magnified to the four corners of the image. As shown in Table 1, Table 1 presents the comparison results of each indicator.

[0062] Table 1

[0063]

[0064] In tests on the entire dataset, the method of this invention achieved a normalized intersection-over-union ratio of 0.7835, an F1 score of 0.8785, a detection rate of 0.9904, and a resolution of 1.3484 × 10⁻⁶. -5 The ALCNet method achieved a false alarm rate of 0.6551, a normalized crossover ratio of 0.7916, an F1 score of 0.9423, and a detection rate of 2.7501 × 10⁻⁶. -5The false alarm rates are shown in Table 1. Higher normalized intersection-over-union ratio (NOR), F1 score, and detection rate result in lower false alarm rates and better target detection performance. Compared to ALCNet, our method improves the NOR by 19.6%, F1 score by 10.9%, and detection rate by 5.1%, while reducing the false alarm rate by 50.9%.

[0065] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0066] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. An infrared small target detection method based on a multi-scale mutual attention mechanism, characterized by: This method is implemented by the following steps: Step 1: Obtain the infrared small target detection dataset; Step 2: Perform data augmentation on the dataset detected in Step 1 to obtain the processed dataset, and divide the dataset into training set, validation set and test set; Step 3: Establish a target detection network model based on multi-faceted feature extraction and mutual attention mechanism; the network model includes a multi-shape receptive field extraction module for feature extraction, a spatial-channel-global-local joint re-extraction module for feature enhancement and re-extraction, and a multi-scale mutual attention mechanism module for feature fusion; input the training set into the network model for training; Step 4: Load the validation set into the network model, calculate the loss function, backpropagate, update the learning rate, and save the network model parameters; Step 5: Load the validation set and the network model parameters saved in Step 4 into the network model, obtain the detection results, and validate the network model; after multiple rounds of training, obtain the optimal model parameters. Step 6: Load the test set and optimal model parameters into the network model to obtain the results of infrared small target detection.

2. The infrared small target detection method based on a multi-scale mutual attention mechanism according to claim 1, characterized in that: The multi-shape receptive field extraction module will use a combination of Res2Net network structure and multi-rate dilated convolution to construct a dynamic receptive field adaptive adjustment mechanism, thereby realizing the extraction of multi-shape contours and multi-scale size features of infrared small targets.

3. The infrared small target detection method based on a multi-scale mutual attention mechanism according to claim 2, characterized in that: The multi-shape receptive field extraction module splits the feature map into four image blocks for differential feature extraction; The second image patch is subjected to deformable dilated convolution, normalization layer, and ReLU activation function operation, and then adjusted with learnable parameters; the result is fused with the third image patch, and then subjected to deformable dilated convolution, normalization, ReLU activation function operation, and then adjusted with learnable parameters. The result is fused with the fourth image patch, and then subjected to deformable dilated convolution, normalization, ReLU activation function operation, and adjusted with learnable parameters; the three processed image patches are then stitched together with the first image patch. Then, it is fused with the feature map before splitting, and then subjected to convolution, batch normalization and ReLU activation function operations. Channel attention mechanism and spatial attention mechanism are used for enhancement to obtain the feature-extracted image as intermediate features.

4. The infrared small target detection method based on a multi-scale mutual attention mechanism according to claim 1, characterized in that: The spatial-channel-global-local joint re-extraction module includes: a local-spatial local-channel module, a global channel extraction module, and a local channel extraction module; The intermediate features are input into the re-extraction module and split into two image blocks. The two image blocks are processed by the local-spatial-local-channel module, the global channel extraction module, and the local channel extraction module, respectively, which output two local features. The two local features are then concatenated and fused with the intermediate features before splitting. The spatial-channel-global-local joint re-extraction module outputs multiple local channel images.

5. The infrared small target detection method based on a multi-scale mutual attention mechanism according to claim 4, characterized in that: The local-spatial-local-channel module is used to extract width and height features from image blocks. Specifically, it performs convolution, batch normalization, and ReLU activation on the segmented image blocks, followed by convolution with a kernel height of 1 and a width of 3, batch normalization, and ReLU activation, then dilated convolution with a kernel height of 1, a width of 3, and a dilation factor of 2, batch normalization, and ReLU activation, and finally convolution with a kernel of 1, batch normalization, and ReLU activation, followed by a sigmoid normalization function. The resulting data is multiplied by the segmented image blocks and then added to complete the residual connection. Then, the output is input into the global channel extraction module, which performs averaging and maximizing operations on the channel dimensions. The results of the averaging and maximizing operations are then concatenated, followed by convolution with a kernel height of 1 and a width of 3, batch normalization, and ReLU activation function operations. After passing through a sigmoid normalization function, the result is multiplied by the image patch input to the global channel extraction module and then added to complete the residual connection. Finally, the image patch output by the global channel extraction module is input into the local channel extraction module. The local channel extraction module performs average downsampling to 1×1 size in the spatial dimension, then adjusts the overall dimension, and then performs one-dimensional convolution with a kernel of 3, batch normalization, and ReLU activation function operation. After adjusting the overall dimension again, it passes through a sigmoid normalization function, multiplies with the image patch input to the local channel extraction module, and then adds it to complete the residual connection.

6. The infrared small target detection method based on a multi-scale mutual attention mechanism according to claim 4, characterized in that: The multiple local channel images output by the spatial-channel-global-local joint re-extraction module are input into the multi-scale mutual attention mechanism module. The multi-scale mutual attention mechanism module downsamples the multi-scale features to the same size, rearranges them, and performs matrix multiplication to obtain the feature fusion matrix. The feature fusion matrix is ​​multiplied with the multi-scale features to obtain the final multi-scale feature layer. After passing through the channel attention mechanism and the spatial attention mechanism, upsampling is performed to obtain the multi-scale feature fusion features.

Images