Infrared small target detection method and system based on scale perception and joint attention

Abstract

The invention provides an infrared small target detection method and system based on scale perception and joint attention, and the method comprises the steps: carrying out the preliminary feature extraction of an infrared image through a basic feature extraction module, and obtaining a primary feature graph; performing multi-level feature extraction on the primary feature map through a multi-scale sensing module to obtain a multi-scale feature map; fusing the primary feature map and the multi-scale feature map through a full semantic aggregation coding module to obtain a coded feature map; inputting the coded feature map and the multi-scale feature map into a cross-semantic guidance decoding module for reconstruction decoding to obtain a decoded feature map; inputting the decoded feature map into a prediction head to obtain a prediction image; and obtaining a final detection result based on the predicted image. According to the method, the model structure is effectively optimized through the two mechanisms of scale perception and joint attention, the model parameters and the calculation load capacity are remarkably reduced, and the high-efficiency and light-weight detection capability is realized while the excellent performance is kept.

Description

Technical Field

[0001] The present invention relates to the technical field of computer vision and target detection, and in particular to a method and system for detecting small infrared targets based on scale perception and joint attention. Background Art

[0002] Infrared small target detection is a key technology in modern image processing and is widely used in intelligent robots, autonomous driving, natural disaster warning, aerospace and other fields. However, due to the long-range imaging mechanism, the captured infrared small target images usually have the characteristics of low contrast, small size, and severe noise interference, which poses a major challenge to the accurate detection of small targets in complex backgrounds. Existing methods to address these challenges are generally divided into traditional methods and deep learning-based methods. Traditional methods measure the discontinuity between the target and the background and obtain prior knowledge through target feature analysis. Various methods have been proposed, including filter-based methods, human visual system-based methods, and low-rank-based methods. Although traditional methods have achieved some success in specific environments, they rely heavily on manual feature engineering and extensive prior knowledge, resulting in limited robustness in complex real-world scenarios.

[0003] Due to the lack of texture and structural information of targets in infrared images, detectors must rely on contextual information to distinguish small targets from complex backgrounds with similar shapes. However, these methods are limited by unreasonable network depth, which restricts the receptive field and hinders the capture of comprehensive contextual information. In addition, the limited receptive field weakens the effective utilization of multi-scale features in various detection stages and increases the risk of false positives. Secondly, the feature utilization between different network layers is still insufficient. Although high-level features contain rich semantic information, downsampling inevitably compromises target resolution and may cause information loss. To alleviate this problem, previous studies introduced feature fusion modules to integrate high-resolution information from low-level features and enhance high-level feature detail recovery. However, low-level features are susceptible to significant noise interference while preserving spatial details. Direct fusion of low-level and high-level features introduces noise and reduces detection accuracy. Summary of the invention

[0004] In view of the above situation, the main purpose of the present invention is to propose an infrared small target detection method and system based on scale perception and joint attention to solve the above technical problems.

[0005] The present invention proposes an infrared small target detection method based on scale perception and joint attention, and the method comprises the following steps: Step 1: Perform preliminary feature extraction on the infrared image through the basic feature extraction module to obtain a primary feature map; Step 2: Perform multi-level feature extraction and channel splicing operations on the primary feature map to obtain a spliced ​​multi-level feature extraction feature map, and then process the spliced ​​multi-level feature extraction feature map through a coordinate attention module to obtain a multi-scale feature map; Step 3: Process the primary feature map and the multi-scale feature map through three sets of projection convolution layers to generate the query matrix, key matrix and value matrix of the primary feature map and the multi-scale feature map respectively; Obtaining an attention refinement feature map of the primary feature map and an attention refinement feature map of the multi-scale feature map based on the query matrix, the key matrix, and the value matrix of the primary feature map and the multi-scale feature map; Obtaining a semantic difference feature map of the primary feature map and a semantic difference feature map of the multi-scale feature map based on the query matrix, the key matrix and the value matrix of the primary feature map and the multi-scale feature map; Performing convolution processing on the attention refinement feature map of the primary feature map, the attention refinement feature map of the multi-scale feature map, the semantic difference feature map of the primary feature map, and the semantic difference feature map of the multi-scale feature map to obtain an encoding feature map; Step 4: Input the encoded feature map and the multi-scale feature map into the cross-semantic guided decoding module for reconstruction and decoding to obtain a decoded feature map; Step 5: Input the decoded feature map into the prediction head to obtain the predicted image; The final detection result is obtained based on the predicted image.

[0006] The present invention also proposes an infrared small target detection system based on scale perception and joint attention, the system comprising: Feature extraction module for: Perform preliminary feature extraction on the infrared image to obtain a primary feature map; Scale-aware module for: Perform multi-level feature extraction on the primary feature map to obtain a multi-scale feature map; Full semantic aggregation encoding module, used for: The primary feature map and the multi-scale feature map are processed through three sets of projection convolution layers to generate the query matrix, key matrix and value matrix of the primary feature map and the multi-scale feature map respectively; Obtaining an attention refinement feature map of the primary feature map and an attention refinement feature map of the multi-scale feature map based on the query matrix, the key matrix, and the value matrix of the primary feature map and the multi-scale feature map; Obtaining a semantic difference feature map of the primary feature map and a semantic difference feature map of the multi-scale feature map based on the query matrix, the key matrix and the value matrix of the primary feature map and the multi-scale feature map; Performing convolution processing on the attention refinement feature map of the primary feature map, the attention refinement feature map of the multi-scale feature map, the semantic difference feature map of the primary feature map, and the semantic difference feature map of the multi-scale feature map to obtain an encoding feature map; Cross-semantically guided decoding module for: Reconstruct and decode the encoded feature map and the multi-scale feature map to obtain a decoded feature map; Image prediction module, used for: Input the decoded feature map into the prediction head to obtain the predicted image; The final detection result is obtained based on the predicted image.

[0007] Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention significantly reduces model parameters and calculation amount by optimizing the model structure, and realizes efficient and lightweight detection capability while maintaining excellent performance, which is suitable for practical applications; 2. The present invention designs a scale perception module to expand the receptive field, so that richer context information can be extracted, thereby significantly improving the detection accuracy and reliability of small infrared targets in complex backgrounds; 3. The present invention designs a full semantic aggregation coding module, which effectively strengthens the feature expression of the target area, suppresses background noise interference, improves the separation effect of the target and the background, and adapts to complex scenes; 4. The present invention designs a cross-semantic guided decoding module, which retains detailed information while refining and reconstructing the target structure, ensuring the integrity and accuracy of the detection results.

[0008] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description or learned through embodiments of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 This is a flow chart of the infrared small target detection method based on scale perception and joint attention proposed by the present invention; Figure 2 This is a schematic diagram of the full semantic aggregation coding module of the infrared small target detection method based on scale perception and joint attention proposed in the present invention; Figure 3 This is a schematic diagram of the cross-semantic guided decoding module of the infrared small target detection method based on scale perception and joint attention proposed in the present invention; Figure 4 Schematic diagram of the overall framework of the infrared small target detection system based on scale perception and joint attention proposed in the present invention. DETAILED DESCRIPTION

[0010] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0011] These and other aspects of the embodiments of the present invention will be apparent with reference to the following description and accompanying drawings. In these descriptions and accompanying drawings, some specific implementations of the embodiments of the present invention are specifically disclosed to represent some ways of implementing the principles of the embodiments of the present invention, but it should be understood that the scope of the embodiments of the present invention is not limited thereto.

[0012] See also Figure 1 The embodiment of the present invention proposes an infrared small target detection method based on scale perception and joint attention, which includes the following steps: Step 1: Perform preliminary feature extraction on the infrared image through the basic feature extraction module to obtain a primary feature map; In step 1, the basic feature extraction module is used to perform preliminary feature extraction on the infrared image to obtain a primary feature map. The relationship between the corresponding process is: ; in, represents the primary feature map, It means that it has been processed by the ReLU activation function. It means that it has been processed by batch normalization layer. It means that the convolution layer with a convolution kernel size of 3×3 is processed. Represents an infrared image.

[0013] Step 2: Perform multi-level feature extraction and channel splicing operations on the primary feature map to obtain a spliced ​​multi-level feature extraction feature map, and then process the spliced ​​multi-level feature extraction feature map through a coordinate attention module to obtain a multi-scale feature map; In step 2, multi-level feature extraction and channel splicing operations are performed on the primary feature map to obtain a spliced ​​multi-level feature extraction feature map, and then the spliced ​​multi-level feature extraction feature map is processed by the coordinate attention module to obtain a multi-scale feature map: ; in, represents a multi-scale feature map, It means that it has been processed by the coordinate attention module. , , and They represent the convolution layers with kernel sizes of 1×1, 5×5, 7×7, and 9×9, respectively. Indicates channel splicing operation; The working principle of the coordinate attention module is as follows: Use the two spatial ranges of the pooling kernel to encode each channel of the input of the coordinate attention module along the horizontal coordinate and the vertical coordinate respectively, and obtain the encoding result in height. The relationship between the corresponding process is: ; in, Indicates Channels at height The encoding result on represents the width of the input image, The index representing the width, represents a specific coordinate in the input image, Indicates Channels at height and width The encoding result at Use the two spatial ranges of the pooling kernel to encode each channel of the input of the coordinate attention module along the horizontal coordinate and the vertical coordinate respectively, and obtain the encoding result in width. The relationship between the corresponding process is: ; in, Indicates Channels in width The encoding result on represents the height of the input image, The index representing the height, represents a specific coordinate in the input image, Indicates Channels at height and width The encoding result at ; The encoding results in height and width are aggregated along two spatial directions to generate a pair of direction-aware feature maps, and then the direction-aware feature maps are connected through convolution to obtain the intermediate feature map. The relationship between the corresponding process is: ; in, represents the intermediate feature map, Indicates the encoding result in height, Indicates the encoding result in width; The intermediate feature map is divided into two tensors along the spatial dimension, and then two convolutions are used to transform the two tensors into attention weights in two directions in the horizontal and vertical directions respectively. The relationship between the corresponding processes is: ; in, represents the attention weight in the vertical direction, represents the attention weight in the horizontal direction, Indicates that it has been processed by the Sigmoid activation function. represents the intermediate feature map in the vertical direction, Represents the intermediate feature map in the horizontal direction; Finally, the input is multiplied by the attention weights in the two directions to obtain the output of the coordinate attention module. The corresponding relationship is: ; in, Represents the output of the coordinate attention module.

[0014] Step 3: Process the primary feature map and the multi-scale feature map through three sets of projection convolution layers to generate the query matrix, key matrix and value matrix of the primary feature map and the multi-scale feature map respectively; Obtaining an attention refinement feature map of the primary feature map and an attention refinement feature map of the multi-scale feature map based on the query matrix, the key matrix, and the value matrix of the primary feature map and the multi-scale feature map; Obtaining a semantic difference feature map of the primary feature map and a semantic difference feature map of the multi-scale feature map based on the query matrix, the key matrix and the value matrix of the primary feature map and the multi-scale feature map; Performing convolution processing on the attention refinement feature map of the primary feature map, the attention refinement feature map of the multi-scale feature map, the semantic difference feature map of the primary feature map, and the semantic difference feature map of the multi-scale feature map to obtain an encoding feature map; See also Figure 2 In step 3, the attention refinement feature map of the primary feature map and the attention refinement feature map of the multi-scale feature map are obtained based on the query matrix, key matrix and value matrix of the primary feature map and the multi-scale feature map. The specific steps are as follows: Integrate the value matrix of the primary feature map and the multi-scale feature map to obtain the cross-semantic value. The relationship between the corresponding process is: ; in, Represents a cross-semantic value, represents the value matrix of the primary feature map, A matrix of values ​​representing multi-scale feature maps; The global attention map of the primary feature map is obtained by using the query matrix of the primary feature map and the key matrix of the primary feature map. The global attention map of the multi-scale feature map is obtained by using the query matrix of the multi-scale feature map and the key matrix of the multi-scale feature map. The relationship between the corresponding processes is: ; in, represents the global attention map of the primary feature map, represents the global attention map of multi-scale feature maps, represents the query matrix of the primary feature map, represents the transposed matrix of the key matrix of the primary feature map, represents the global attention map of multi-scale feature maps, The transposed matrix of the key matrix representing the multi-scale feature map; The global attention map of the primary feature map and the global attention map of the multi-scale feature map are refined by using the cross-semantic value to obtain the attention refinement feature map of the primary feature map and the attention refinement feature map of the multi-scale feature map. The relationship between the corresponding processes is: ; in, represents the attention-refined feature map of the primary feature map, Attention-refined feature maps representing multi-scale feature maps; The semantic difference feature map of the primary feature map and the semantic difference feature map of the multi-scale feature map are obtained based on the query matrix, key matrix and value matrix of the primary feature map and the multi-scale feature map. The specific steps are as follows: The query matrix of the multi-scale feature map and the key matrix of the primary feature map are used to obtain the cross-semantic attention map of the primary feature map. The query matrix of the primary feature map and the key matrix of the multi-scale feature map are used to obtain the cross-semantic attention map of the multi-scale feature map. The relationship between the corresponding processes is: ; in, represents the cross semantic attention map of the primary feature map, Cross-semantic attention map representing multi-scale feature maps; Based on the value matrix of the primary feature map and the cross semantic attention map of the primary feature map, the semantic difference feature map of the primary feature map is obtained. Based on the value matrix of the multi-scale feature map and the cross semantic attention map of the multi-scale feature map, the semantic difference feature map of the multi-scale feature map is obtained. The relationship between the corresponding processes is: ; in, The semantic difference feature map representing the primary feature map, A semantic difference feature map representing a multi-scale feature map; The attention refinement feature map of the primary feature map, the attention refinement feature map of the multi-scale feature map, the semantic difference feature map of the primary feature map, and the semantic difference feature map of the multi-scale feature map are convolved to obtain the encoding feature map. The relationship between the corresponding processes is: ; in, Represents the encoded feature map.

[0015] Step 4: Input the encoded feature map and the multi-scale feature map into the cross-semantic guided decoding module for reconstruction and decoding to obtain a decoded feature map; See also Figure 3 In step 4, the encoded feature map and the multi-scale feature map are input into the cross-semantic guided decoding module for reconstruction and decoding to obtain the decoded feature map. The specific steps are as follows: The encoding feature map and the multi-scale feature map are processed by convolution layer and activation function to generate the spatial weight matrix of the encoding feature map and the spatial weight matrix of the multi-scale feature map respectively. The relationship between the corresponding processes is: ; in, represents the spatial weight matrix of the encoded feature map, A spatial weight matrix representing a multi-scale feature map; The spatial weight matrix of the multi-scale feature map is multiplied element by element with the encoding feature map at the feature level to obtain the encoding feature map with spatial information enhancement. The spatial weight matrix of the encoding feature map is multiplied element by element with the multi-scale feature map at the feature level to obtain the multi-scale feature map with spatial information enhancement. The corresponding process has the following relationship: ; in, Represents the encoded feature map with enhanced spatial information, Represents spatial information enhanced multi-scale feature map; The spatial information enhanced encoding feature map and the spatial information enhanced multi-scale feature map are subjected to channel splicing operations, and are processed by the convolution layer, the coordinate attention module and the activation function in sequence to generate a fusion weight matrix. The relationship between the corresponding processes is: ; in, represents the fusion weight matrix; The fusion matrix is ​​applied to the encoding feature map and the multi-scale feature map for dynamic reweighting, and then refined through the convolution layer to obtain the decoding features. The corresponding process has the following relationship: ; in, Represents the decoded feature map.

[0016] Step 5: Input the decoded feature map into the prediction head to obtain the predicted image; Obtain the final detection result based on the predicted image; The decoded feature map is input into the prediction head to obtain the predicted image. The relationship between the corresponding process is: ; in, represents the predicted image, Represents upsampling via bilinear interpolation.

[0017] See also Figure 4 The embodiment of the present invention further provides an infrared small target detection system based on scale perception and joint attention, the system comprising: Feature extraction module for: Perform preliminary feature extraction on the infrared image to obtain a primary feature map; Scale-aware module for: Perform multi-level feature extraction on the primary feature map to obtain a multi-scale feature map; Full semantic aggregation encoding module, used for: The primary feature map and the multi-scale feature map are processed through three sets of projection convolution layers to generate the query matrix, key matrix and value matrix of the primary feature map and the multi-scale feature map respectively; Obtaining an attention refinement feature map of the primary feature map and an attention refinement feature map of the multi-scale feature map based on the query matrix, the key matrix, and the value matrix of the primary feature map and the multi-scale feature map; Obtaining a semantic difference feature map of the primary feature map and a semantic difference feature map of the multi-scale feature map based on the query matrix, the key matrix and the value matrix of the primary feature map and the multi-scale feature map; Performing convolution processing on the attention refinement feature map of the primary feature map, the attention refinement feature map of the multi-scale feature map, the semantic difference feature map of the primary feature map, and the semantic difference feature map of the multi-scale feature map to obtain an encoding feature map; Cross-semantically guided decoding module for: Reconstruct and decode the encoded feature map and the multi-scale feature map to obtain a decoded feature map; Image prediction module, used for: Input the decoded feature map into the prediction head to obtain the predicted image; The final detection result is obtained based on the predicted image.

[0018] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0019] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0020] The above-mentioned embodiments only express several implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A method for infrared small target detection based on scale perception and joint attention, characterized in that: The method comprises the following steps: Step 1: Perform preliminary feature extraction on the infrared image through the basic feature extraction module to obtain a primary feature map; Step 2: Perform multi-level feature extraction and channel splicing operations on the primary feature map to obtain a spliced ​​multi-level feature extraction feature map, and then process the spliced ​​multi-level feature extraction feature map through a coordinate attention module to obtain a multi-scale feature map; Step 3: Process the primary feature map and the multi-scale feature map through three sets of projection convolution layers to generate the query matrix, key matrix and value matrix of the primary feature map and the multi-scale feature map respectively; Obtaining an attention refinement feature map of the primary feature map and an attention refinement feature map of the multi-scale feature map based on the query matrix, the key matrix, and the value matrix of the primary feature map and the multi-scale feature map; Obtaining a semantic difference feature map of the primary feature map and a semantic difference feature map of the multi-scale feature map based on the query matrix, the key matrix and the value matrix of the primary feature map and the multi-scale feature map; Performing convolution processing on the attention refinement feature map of the primary feature map, the attention refinement feature map of the multi-scale feature map, the semantic difference feature map of the primary feature map, and the semantic difference feature map of the multi-scale feature map to obtain an encoding feature map; Step 4: Input the encoded feature map and the multi-scale feature map into the cross-semantic guided decoding module for reconstruction and decoding to obtain a decoded feature map; Step 5: Input the decoded feature map into the prediction head to obtain the predicted image; The final detection result is obtained based on the predicted image.

2. The infrared small target detection method based on scale perception and joint attention according to claim 1 is characterized in that: In step 1, the infrared image is subjected to preliminary feature extraction by the basic feature extraction module to obtain a primary feature map. The relationship between the corresponding process is: ； in, represents the primary feature map, It means that it has been processed by the ReLU activation function. It means that it has been processed by batch normalization layer. It means that the convolution layer with a convolution kernel size of 3×3 is processed. Represents an infrared image.

3. The infrared small target detection method based on scale perception and joint attention according to claim 2 is characterized in that: In step 2, multi-level feature extraction and channel splicing operations are performed on the primary feature map to obtain a spliced ​​multi-level feature extraction feature map, and then the spliced ​​multi-level feature extraction feature map is processed by the coordinate attention module to obtain a multi-scale feature map. The relationship between the corresponding process is: ； in, represents a multi-scale feature map, It means that it has been processed by the coordinate attention module. , , and They represent the convolution layers with kernel sizes of 1×1, 5×5, 7×7, and 9×9, respectively. Indicates channel splicing operation; The working principle of the coordinate attention module is as follows: Use the two spatial ranges of the pooling kernel to encode each channel of the input of the coordinate attention module along the horizontal coordinate and the vertical coordinate respectively, and obtain the encoding result in height. The relationship between the corresponding process is: ； in, Indicates Channels at height The encoding result on represents the width of the input image, The index representing the width, represents a specific coordinate in the input image, Indicates Channels at height and width The encoding result at Use the two spatial ranges of the pooling kernel to encode each channel of the input of the coordinate attention module along the horizontal coordinate and the vertical coordinate respectively, and obtain the encoding result in width. The relationship between the corresponding process is: ； in, Indicates Channels in width The encoding result on represents the height of the input image, The index representing the height, represents a specific coordinate in the input image, Indicates Channels at height and width The encoding result at The encoding results in height and width are aggregated along two spatial directions to generate a pair of direction-aware feature maps, and then the direction-aware feature maps are connected through convolution to obtain the intermediate feature map. The relationship between the corresponding process is: ； in, represents the intermediate feature map, Indicates the encoding result in height, Indicates the encoding result in width; The intermediate feature map is divided into two tensors along the spatial dimension, and then two convolutions are used to transform the two tensors into attention weights in two directions in the horizontal and vertical directions respectively. The relationship between the corresponding processes is: ； in, represents the attention weight in the vertical direction, represents the attention weight in the horizontal direction, Indicates that it has been processed by the Sigmoid activation function. represents the intermediate feature map in the vertical direction, Represents the intermediate feature map in the horizontal direction; Multiply the input by the attention weights in the two directions to get the output of the coordinate attention module. The corresponding relationship is: ； in, Represents the output of the coordinate attention module.

4. The infrared small target detection method based on scale perception and joint attention according to claim 3 is characterized in that: In step 3, the attention refinement feature map of the primary feature map and the attention refinement feature map of the multi-scale feature map are obtained based on the query matrix, key matrix and value matrix of the primary feature map and the multi-scale feature map. The specific steps are as follows: Integrate the value matrix of the primary feature map and the multi-scale feature map to obtain the cross-semantic value. The relationship between the corresponding process is: ； in, Represents a cross-semantic value, A matrix of values ​​representing the primary feature map, A matrix of values ​​representing multi-scale feature maps; The global attention map of the primary feature map is obtained by using the query matrix of the primary feature map and the key matrix of the primary feature map. The global attention map of the multi-scale feature map is obtained by using the query matrix of the multi-scale feature map and the key matrix of the multi-scale feature map. The relationship between the corresponding processes is: ； in, represents the global attention map of the primary feature map, represents the global attention map of multi-scale feature maps, represents the query matrix of the primary feature map, represents the transposed matrix of the key matrix of the primary feature map, represents the global attention map of multi-scale feature maps, The transposed matrix of the key matrix representing the multi-scale feature map; The global attention map of the primary feature map and the global attention map of the multi-scale feature map are refined by using the cross-semantic value to obtain the attention refinement feature map of the primary feature map and the attention refinement feature map of the multi-scale feature map. The relationship between the corresponding processes is: ； in, represents the attention-refined feature map of the primary feature map, Representing the attention-refined feature map of multi-scale feature maps.

5. The infrared small target detection method based on scale perception and joint attention according to claim 4 is characterized in that: In step 3, a semantic difference feature map of the primary feature map and a semantic difference feature map of the multi-scale feature map are obtained based on the query matrix, key matrix and value matrix of the primary feature map and the multi-scale feature map. The specific steps are as follows: The query matrix of the multi-scale feature map and the key matrix of the primary feature map are used to obtain the cross-semantic attention map of the primary feature map. The query matrix of the primary feature map and the key matrix of the multi-scale feature map are used to obtain the cross-semantic attention map of the multi-scale feature map. The relationship between the corresponding processes is: ； in, represents the cross semantic attention map of the primary feature map, Cross-semantic attention map representing multi-scale feature maps; Based on the value matrix of the primary feature map and the cross semantic attention map of the primary feature map, the semantic difference feature map of the primary feature map is obtained. Based on the value matrix of the multi-scale feature map and the cross semantic attention map of the multi-scale feature map, the semantic difference feature map of the multi-scale feature map is obtained. The relationship between the corresponding processes is: ； in, The semantic difference feature map representing the primary feature map, Semantic difference feature map representing multi-scale feature maps.

6. The infrared small target detection method based on scale perception and joint attention according to claim 5 is characterized in that: In step 3, the attention refinement feature map of the primary feature map, the attention refinement feature map of the multi-scale feature map, the semantic difference feature map of the primary feature map, and the semantic difference feature map of the multi-scale feature map are convolved to obtain the encoding feature map. The relationship between the corresponding processes is: ； in, Represents the encoded feature map.

7. The infrared small target detection method based on scale perception and joint attention according to claim 6 is characterized in that: In step 4, the encoded feature map and the multi-scale feature map are input into the cross-semantic guided decoding module for reconstruction and decoding to obtain a decoded feature map. The specific steps are as follows: The encoding feature map and the multi-scale feature map are processed by convolution layer and activation function to generate the spatial weight matrix of the encoding feature map and the spatial weight matrix of the multi-scale feature map respectively. The relationship between the corresponding processes is: ； in, represents the spatial weight matrix of the encoded feature map, A spatial weight matrix representing a multi-scale feature map; The spatial weight matrix of the multi-scale feature map is multiplied element by element with the encoding feature map at the feature level to obtain the encoding feature map with spatial information enhancement. The spatial weight matrix of the encoding feature map is multiplied element by element with the multi-scale feature map at the feature level to obtain the multi-scale feature map with spatial information enhancement. The corresponding process has the following relationship: ； in, Represents the encoded feature map with enhanced spatial information, Represents spatial information enhanced multi-scale feature map; The spatial information enhanced encoding feature map and the spatial information enhanced multi-scale feature map are subjected to channel splicing operations, and are processed by the convolution layer, the coordinate attention module and the activation function in sequence to generate a fusion weight matrix. The relationship between the corresponding processes is: ； in, represents the fusion weight matrix; The fusion matrix is ​​applied to the encoding feature map and the multi-scale feature map for dynamic reweighting, and then refined through the convolution layer to obtain the decoding features. The corresponding process has the following relationship: ； in, Represents the decoded feature map.

8. The infrared small target detection method based on scale perception and joint attention according to claim 7 is characterized in that: In step 5, the decoded feature map is input into the prediction head to obtain a predicted image. The relationship between the corresponding process is: ； in, represents the predicted image, Represents upsampling via bilinear interpolation.

9. An infrared small target detection system based on scale perception and joint attention, characterized in that: The system applies any one of the infrared small target detection methods based on scale perception and joint attention as claimed in claims 1 to 8 above, and the system comprises: Feature extraction module for: Perform preliminary feature extraction on the infrared image to obtain a primary feature map; Scale-aware module for: Perform multi-level feature extraction on the primary feature map to obtain a multi-scale feature map; Full semantic aggregation encoding module, used for: Process the primary feature map and the multi-scale feature map through three sets of projection convolution layers to generate the query matrix, key matrix and value matrix of the primary feature map and the multi-scale feature map respectively; Obtaining an attention refinement feature map of the primary feature map and an attention refinement feature map of the multi-scale feature map based on the query matrix, the key matrix, and the value matrix of the primary feature map and the multi-scale feature map; Obtaining a semantic difference feature map of the primary feature map and a semantic difference feature map of the multi-scale feature map based on the query matrix, the key matrix and the value matrix of the primary feature map and the multi-scale feature map; Performing convolution processing on the attention refinement feature map of the primary feature map, the attention refinement feature map of the multi-scale feature map, the semantic difference feature map of the primary feature map, and the semantic difference feature map of the multi-scale feature map to obtain an encoding feature map; Cross-semantically guided decoding module for: Reconstruct and decode the encoded feature map and the multi-scale feature map to obtain a decoded feature map; Image prediction module, used for: Input the decoded feature map into the prediction head to obtain the predicted image; The final detection result is obtained based on the predicted image.

Images