Infrared Small Target Detection Method and System Based on Adaptive Partial Self-Attention
By adopting an adaptive partial self-attention infrared small target detection method, the problem of infrared small target recognition in complex backgrounds and low-contrast environments is solved, achieving efficient and accurate infrared small target detection, and adapting to a variety of application scenarios.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-23
- Publication Date
- 2026-04-03
AI Technical Summary
Existing technologies struggle to accurately identify small infrared targets in complex backgrounds and low-contrast environments. Traditional methods suffer from high false negative rates and severe false positives, while CNN methods have shortcomings in long-distance dependency modeling and contextual information capture.
An adaptive partial self-attention infrared small target detection method is adopted. The infrared image is processed by dual residual attention blocks and feature encoders. Self-attention calculation of multi-scale high-level coded feature maps and generation of global context long-range feature maps are performed. Target detection is performed by combining transposed convolution and prediction head.
It can accurately identify small targets in complex backgrounds and low signal-to-noise ratio environments, reduce missed detections and false detections, lower computational costs, adapt to various application scenarios, and meet the real-time detection needs of resource-constrained scenarios.
Smart Images
Figure CN119540539B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of computer vision and target detection, and in particular to an infrared small target detection method and system based on adaptive partial self-attention. Background Technology
[0002] Infrared small target detection is an important research direction in the fields of computer vision and target detection, with wide practical applications in military, security, and rescue scenarios. However, due to the weak signal of infrared small targets in images, they typically exhibit characteristics such as low contrast, small size, and complex backgrounds, making it difficult for traditional detection methods to accurately identify and detect these targets. Current technologies typically employ traditional methods based on filtering, background modeling, and machine learning, or deep learning methods using Convolutional Neural Networks (CNNs), for infrared small target detection. However, these methods still have significant limitations in complex environments.
[0003] Traditional infrared small target detection methods mainly rely on background suppression and filtering techniques, such as high-pass filtering, differential operators, and background prediction, to enhance the signal of small targets. However, these methods often perform poorly in scenarios with complex backgrounds, low contrast, and low signal-to-noise ratios. Furthermore, traditional methods lack the ability to deeply represent target features, making it difficult to adapt to the diverse target characteristics in infrared images, resulting in high false negative rates and serious false positives.
[0004] With the development of deep learning technology, infrared small target detection methods based on convolutional neural networks (CNNs) have attracted widespread attention. CNNs excel in image feature extraction, gradually acquiring rich feature information from low to high layers by constructing multi-layer convolutional and pooling network structures. However, CNN methods are mainly based on local receptive fields and have shortcomings in long-distance dependency modeling and contextual information capture. Infrared small targets are usually small in size and have weak signals, making it difficult to effectively enhance the target signal by simply relying on convolutional operations. In addition, because the feature extraction method of CNNs is fixed, it is difficult to adapt to changes in the target, resulting in limited detection accuracy, especially when the infrared image has strong background interference, making it difficult to accurately identify the target. Summary of the Invention
[0005] In view of the above, the main objective of this invention is to propose an infrared small target detection method and system based on adaptive partial self-attention, so as to solve the above-mentioned technical problems.
[0006] This invention proposes an infrared small target detection method based on adaptive partial self-attention, the method comprising the following steps:
[0007] Step 1: Process the single-frame infrared image sequentially through a dual residual attention block and a feature encoder to obtain the encoded feature map;
[0008] Step 2: Perform max pooling downsampling on the coding feature maps at different levels to obtain coding features. Then, concatenate the coding features along the channel dimension to obtain a multi-scale high-level coding feature map.
[0009] Step 3: Perform channel segmentation on the multi-scale high-level coding feature map to obtain the segmented multi-scale high-level coding feature map. Perform self-attention calculation on the segmented multi-scale high-level coding feature map to obtain the global attention map.
[0010] The global attention map and the multi-scale high-level coding feature map are concatenated along the channel dimension to obtain a fused feature map. The fused feature map is then processed by convolutional layers and multilayer perceptrons to obtain a global context long-range feature map.
[0011] Step 4: Perform upsampling operation based on transpose convolution on the global context long-range feature map and then merge it with the encoded feature map to obtain the decoded feature map of the intermediate layer. Input the decoded feature map of the intermediate layer into the feature decoder to obtain the decoded feature map.
[0012] Step 5: Input the decoded feature map into the prediction head to obtain the latent prediction image, integrate the latent prediction images to obtain the final detection image;
[0013] The target is tracked based on the final detected image.
[0014] This invention also proposes an infrared small target detection system based on adaptive partial self-attention, the system comprising:
[0015] The feature encoding module is used for:
[0016] A single-frame infrared image is processed sequentially through a dual residual attention block and a feature encoder to obtain an encoded feature map;
[0017] Max pooling downsampling is performed on the coding feature maps at different levels to obtain coding features. The coding features are then concatenated along the channel dimension to obtain a multi-scale high-level coding feature map.
[0018] An adaptive partial self-attention module is used for:
[0019] The multi-scale high-level coding feature map is subjected to channel segmentation to obtain the segmented multi-scale high-level coding feature map. Self-attention is then calculated on the segmented multi-scale high-level coding feature map to obtain the global attention map.
[0020] The global attention map and the multi-scale high-level coding feature map are concatenated along the channel dimension to obtain a fused feature map. The fused feature map is then processed by convolutional layers and multilayer perceptrons to obtain a global context long-range feature map.
[0021] The feature decoding module is used for:
[0022] The global context long-range feature map is upsampled based on transpose convolution and then merged with the encoded feature map to obtain the intermediate layer's decoded feature map. The intermediate layer's decoded feature map is then input into the feature decoder to obtain the decoded feature map.
[0023] Image detection module, used for:
[0024] The decoded feature map is input into the prediction head to obtain the latent prediction image. The latent prediction images are then integrated to obtain the final detection image.
[0025] The target is tracked based on the final detected image.
[0026] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0027] 1. This invention enhances the modeling of long-range dependencies of small infrared targets through an adaptive partial self-attention module and a multi-scale feature extraction architecture, enabling the system to accurately identify small targets even in complex backgrounds and low-contrast environments, effectively reducing missed detections and false detections.
[0028] 2. This invention enables the system to achieve efficient real-time detection even in resource-constrained scenarios. By employing an adaptive partial Transformer structure, attention calculation is performed only on features from a subset of channels, significantly reducing computational costs.
[0029] 3. This invention enhances the stability of the system in complex environments such as low signal-to-noise ratio and dynamic background, and adapts to various application scenarios to meet various infrared small target detection needs.
[0030] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by means of embodiments of the invention. Attached Figure Description
[0031] Figure 1 This is a flowchart of the infrared small target detection method based on adaptive partial self-attention proposed in this invention;
[0032] Figure 2 This is a schematic diagram of the dual residual attention block in the infrared small target detection method based on adaptive partial self-attention proposed in this invention;
[0033] Figure 3This is a schematic diagram of the adaptive partial self-attention module of the infrared small target detection method based on adaptive partial self-attention proposed in this invention;
[0034] Figure 4 This is a framework diagram of the infrared small target detection system based on adaptive partial self-attention proposed in this invention. Detailed Implementation
[0035] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.
[0036] These and other aspects of the embodiments of the present invention will become clear from the following description and accompanying drawings. In these descriptions and drawings, some specific embodiments of the present invention are specifically disclosed to illustrate some ways of implementing the principles of the embodiments of the present invention; however, it should be understood that the scope of the embodiments of the present invention is not limited thereto.
[0037] Please see Figure 1 and Figure 2 This invention proposes an infrared small target detection method based on adaptive partial self-attention, which includes the following steps:
[0038] Step 1: Process the single-frame infrared image sequentially through a dual residual attention block and a feature encoder to obtain the encoded feature map;
[0039] In step 1, a single-frame infrared image is sequentially processed by a dual residual attention block and a feature encoder to obtain an encoded feature map. The corresponding relationship in this process is as follows:
[0040] ;
[0041] in, Represents a single frame of infrared image. This indicates that the data has been processed by a convolutional layer with a kernel size of 1×1. This indicates that the data has been processed using a double residual attention block. Indicates the first The encoded feature map of the layer, This indicates that the data has been processed through a max pooling layer with a window size of 2 and a stride of 2. Indicates the first The encoded feature map of the layer, Indicates window size. Indicates the step size. Represents the encoded feature map;
[0042] The working principle of the dual residual attention block is expressed by the following relationship:
[0043] ;
[0044] in, Indicates the first The primary coding feature map of the layer, This indicates that the data has been processed by a convolutional layer with a kernel size of 3×3. This indicates that the process has been performed using the Gaussian activation function. This indicates that the batch normalization layer has been applied. Indicates the first Layer of dual residual attention block input features, Indicates the first Layer-enhanced residual feature map Indicates the first after spatial attention enhancement The feature map obtained after the layer feature map is further enhanced by channel attention is... This indicates that the data has been processed through a multi-layered sensing network. This indicates that it has undergone max pooling. This indicates that the average pooling process has been performed. express Feature maps obtained after spatial attention enhancement This indicates that the data has been processed by a convolutional layer with a kernel size of 7×7. This represents the Sigmoid activation function. This indicates a splicing operation at the channel dimension.
[0045] Step 2: Perform max pooling downsampling on the coding feature maps at different levels to obtain coding features. Then, concatenate the coding features along the channel dimension to obtain a multi-scale high-level coding feature map.
[0046] In step 2, max pooling downsampling is performed on the encoded feature maps at different levels to obtain encoded features. These encoded features are then concatenated along the channel dimension to obtain a multi-scale high-level encoded feature map. The corresponding relationship in this process is as follows:
[0047] ;
[0048] in, Indicates the first The coding features of the layer , , and All represent coding features. This represents a multi-scale high-level coded feature map.
[0049] Step 3: Perform channel segmentation on the multi-scale high-level coding feature map to obtain the segmented multi-scale high-level coding feature map. Perform self-attention calculation on the segmented multi-scale high-level coding feature map to obtain the global attention map.
[0050] The global attention map and the multi-scale high-level coding feature map are concatenated along the channel dimension to obtain a fused feature map. The fused feature map is then processed by convolutional layers and multilayer perceptrons to obtain a global context long-range feature map.
[0051] Please see Figure 3 In step 3, the multi-scale high-level coding feature map is subjected to channel segmentation to obtain the segmented multi-scale high-level coding feature map. Self-attention is then calculated on the segmented multi-scale high-level coding feature map to obtain the global attention map. The process also includes the following sub-steps:
[0052] Adaptive segmentation of multi-scale high-level coded feature maps corresponds to the following relationship:
[0053] ;
[0054] in, This indicates that the process is performed using the SiLU activation function. This represents a multi-scale high-level coding feature map after sequential processing through convolution, batch normalization, and SiLU activation function enhancement. Indicates the channel segmentation factor. This indicates the feature of the multi-head self-attention input after channel segmentation. Features representing the number of remaining channels , , and Both represent channel dimensions. Represents the set of real numbers. Indicates the width of a single frame of infrared image. Indicates the height of a single frame of infrared image;
[0055] After the execution channel is segmented and fed into multi-head self-attention, the features are fed into a projection convolutional layer to generate a query weight matrix, a key weight matrix, and a value weight matrix. The corresponding relationship in this process is as follows:
[0056] ;
[0057] in, , and These represent the query weight matrix, key weight matrix, and value weight matrix, respectively. Indicates channel dimension, Indicates the number of attention heads;
[0058] Based on the query weight matrix, key weight matrix, and value weight matrix, the query matrix, key matrix, and value matrix are obtained. Then, the attention score is calculated using these matrices. The corresponding relationship in this process is as follows:
[0059] ;
[0060] in, Indicates the first Attention score, This indicates that the data has been processed using the Softmax function. Representing dimension, Indicates participation in the The query matrix for calculating attention head scores. Participating in the The transpose of the key matrix used to calculate attention head scores Indicates participation in the The value matrix of attention head scores is calculated. This represents the total score for attention projection;
[0061] Self-attention is calculated on the segmented multi-scale high-level coding feature maps to obtain the global attention map. The corresponding relationship in this process is as follows:
[0062] ;
[0063] in, This represents the position embedding matrix of the multi-scale high-level coded feature map. This represents a primary global attention graph. Represents a global attention graph. This represents the primary global attention map enhanced by residual connections;
[0064] The global attention map and the multi-scale high-level coding feature map are concatenated along the channel dimension to obtain a fused feature map. This fused feature map is then processed through convolutional layers and a multilayer perceptron to obtain the global context long-range feature map. The corresponding relationship in this process is as follows:
[0065] ;
[0066] in, Represents the fused feature map. This indicates that the data has been processed by the Dropout layer. This represents the long-range feature map of the global context.
[0067] Furthermore, in this step, the multi-scale gratitude coding feature map is divided into two parts: one part participates in self-attention calculation, and the other part is concatenated with the global attention map on the channel.
[0068] Step 4: Perform upsampling operation based on transpose convolution on the global context long-range feature map and then merge it with the encoded feature map to obtain the decoded feature map of the intermediate layer. Input the decoded feature map of the intermediate layer into the feature decoder to obtain the decoded feature map.
[0069] In step 4, the global context long-range feature map is upsampled using transposed convolution and then merged with the encoded feature map to obtain the intermediate layer's decoded feature map. This intermediate layer's decoded feature map is then input into the feature decoder to obtain the decoded feature map. The corresponding relationship in this process is as follows:
[0070] ;
[0071] in, Represents the encoded feature map, This indicates an upsampling operation based on transpose convolution. Indicates step size is Convolution kernel is transposed convolution, This indicates that the convolution has been performed using a transposed convolution with a kernel size of 2 and a stride of 2. This indicates the number of layers in the decoder.
[0072] Step 5: Input the decoded feature map into the prediction head to obtain the latent prediction image, integrate the latent prediction images to obtain the final detection image;
[0073] Track the target based on the final detected image;
[0074] In step 5, the decoded feature map is input into the prediction head to obtain the latent prediction image. The latent prediction images are then integrated to obtain the final detection image. The corresponding relationship in this process is as follows:
[0075] ;
[0076] in, Represents the potential prediction image, This indicates that the process has been performed using the Sigmoid activation function. This indicates that the data has undergone upsampling processing using bilinear interpolation. This represents the final detected image.
[0077] Please see Figure 4 This invention also provides an infrared small target detection system based on adaptive partial self-attention, the system comprising:
[0078] The feature encoding module is used for:
[0079] A single-frame infrared image is processed sequentially through a dual residual attention block and a feature encoder to obtain an encoded feature map;
[0080] Max pooling downsampling is performed on the coding feature maps at different levels to obtain coding features. The coding features are then concatenated along the channel dimension to obtain a multi-scale high-level coding feature map.
[0081] An adaptive partial self-attention module is used for:
[0082] The multi-scale high-level coding feature map is subjected to channel segmentation to obtain the segmented multi-scale high-level coding feature map. Self-attention is then calculated on the segmented multi-scale high-level coding feature map to obtain the global attention map.
[0083] The global attention map and the multi-scale high-level coding feature map are concatenated along the channel dimension to obtain a fused feature map. The fused feature map is then processed by convolutional layers and multilayer perceptrons to obtain a global context long-range feature map.
[0084] The feature decoding module is used for:
[0085] The global context long-range feature map is upsampled based on transpose convolution and then merged with the encoded feature map to obtain the intermediate layer's decoded feature map. The intermediate layer's decoded feature map is then input into the feature decoder to obtain the decoded feature map.
[0086] Image detection module, used for:
[0087] The decoded feature map is input into the prediction head to obtain the latent prediction image. The latent prediction images are then integrated to obtain the final detection image.
[0088] The target is tracked based on the final detected image.
[0089] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.
[0090] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.
[0091] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. 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 modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.
Claims
1. A method for detecting small infrared targets based on adaptive partial self-attention, the method being implemented based on a prediction model, the prediction model comprising a feature encoding module, an adaptive partial self-attention module, a feature decoding module, and an image detection module, characterized in that, The method includes the following steps: Step 1: The single-frame infrared image is processed sequentially through a dual residual attention block and a feature encoder to obtain the encoded feature map. The corresponding relationship in this process is as follows: ； in, Represents a single frame of infrared image. This indicates that the data has been processed by a convolutional layer with a kernel size of 1×1. This indicates that the data has been processed using a double residual attention block. Indicates the first The encoded feature map of the layer, This indicates that the data has been processed through a max pooling layer with a window size of 2 and a stride of 2. Indicates the first The encoded feature map of the layer, Indicates window size. Indicates the step size. Represents the encoded feature map; The working principle of the dual residual attention block is expressed by the following relationship: ； in, Indicates the first The primary coding feature map of the layer, This indicates that the data has been processed by a convolutional layer with a kernel size of 3×3. This indicates that the process has been performed using the Gaussian activation function. This indicates that the batch normalization layer has been applied. Indicates the first Layer of dual residual attention block input features, Indicates the first Layer-enhanced residual feature map Indicates the first after spatial attention enhancement The feature map obtained after the layer feature map is further enhanced by channel attention is... This indicates that the data has been processed through a multi-layered sensing network. This indicates that it has undergone max pooling. This indicates that the average pooling process has been performed. express Feature maps obtained after spatial attention enhancement This indicates that the data has been processed by a convolutional layer with a kernel size of 7×7. This represents the Sigmoid activation function. This indicates a splicing operation along the channel dimension; Step 2: Perform max pooling downsampling on the coding feature maps at different levels to obtain coding features. Then, concatenate the coding features along the channel dimension to obtain a multi-scale high-level coding feature map. Step 3: Perform channel segmentation on the multi-scale high-level coding feature map to obtain the segmented multi-scale high-level coding feature map. Perform self-attention calculation on the segmented multi-scale high-level coding feature map to obtain the global attention map. The global attention map and the multi-scale high-level coding feature map are concatenated along the channel dimension to obtain a fused feature map. The fused feature map is then processed by convolutional layers and multilayer perceptrons to obtain a global context long-range feature map. Step 4: Perform upsampling operation based on transpose convolution on the global context long-range feature map and then merge it with the encoded feature map to obtain the decoded feature map of the intermediate layer. Input the decoded feature map of the intermediate layer into the feature decoder to obtain the decoded feature map. Step 5: Input the decoded feature map into the prediction head to obtain the latent prediction image, integrate the latent prediction images to obtain the final detection image; The target is tracked based on the final detected image.
2. The infrared small target detection method based on adaptive partial self-attention according to claim 1, characterized in that, In step 2, max pooling downsampling is performed on the encoded feature maps at different levels to obtain encoded features. These encoded features are then concatenated along the channel dimension to obtain a multi-scale high-level encoded feature map. The corresponding relationship in this process is as follows: ； in, Indicates the first The coding features of the layer , , and All represent coding features. This represents a multi-scale high-level coded feature map.
3. The infrared small target detection method based on adaptive partial self-attention according to claim 2, characterized in that, In step 3, the multi-scale high-level coding feature map is subjected to channel segmentation to obtain a segmented multi-scale high-level coding feature map. Self-attention is then calculated on the segmented multi-scale high-level coding feature map to obtain a global attention map. The process also includes the following sub-steps: Adaptive segmentation of multi-scale high-level coded feature maps corresponds to the following relationship: ； in, This indicates that the process is performed using the SiLU activation function. This represents a multi-scale high-level coding feature map after sequential processing through convolution, batch normalization, and SiLU activation function enhancement. Indicates the channel segmentation factor. This indicates the feature of the multi-head self-attention input after channel segmentation. Features representing the number of remaining channels , , and Both represent channel dimensions. Represents the set of real numbers. Indicates the width of a single frame of infrared image. Indicates the height of a single frame of infrared image; After the execution channel is segmented and fed into multi-head self-attention, the features are fed into a projection convolutional layer to generate a query weight matrix, a key weight matrix, and a value weight matrix. The corresponding relationship in this process is as follows: ； in, , and These represent the query weight matrix, key weight matrix, and value weight matrix, respectively. Indicates channel dimension, Indicates the number of attention heads; Based on the query weight matrix, key weight matrix, and value weight matrix, the query matrix, key matrix, and value matrix are obtained. Then, the attention score is calculated using these matrices. The corresponding relationship in this process is as follows: ； in, Indicates the first Attention score, This indicates that the data has been processed using the Softmax function. Representing dimension, Indicates participation in the The query matrix for calculating attention head scores. Participating in the The transpose of the key matrix used to calculate attention head scores Indicates participation in the The value matrix of attention head scores is calculated. This represents the total score for attention projection; Self-attention is calculated on the segmented multi-scale high-level coding feature maps to obtain the global attention map. The corresponding relationship in this process is as follows: ； in, This represents the position embedding matrix of the multi-scale high-level coded feature map. This represents a primary global attention graph. Represents a global attention graph. This represents the primary global attention graph enhanced by residual connections.
4. The infrared small target detection method based on adaptive partial self-attention according to claim 3, characterized in that, In step 3, the global attention map and the multi-scale high-level coding feature map are concatenated along the channel dimension to obtain a fused feature map. This fused feature map is then processed by convolutional layers and a multilayer perceptron to obtain a global context long-range feature map. The corresponding relationship in this process is as follows: ； in, Represents the fused feature map. This indicates that the data has been processed by the Dropout layer. This represents the long-range feature map of the global context.
5. The infrared small target detection method based on adaptive partial self-attention according to claim 4, characterized in that, In step 4, the global context long-range feature map is upsampled using transposed convolution and then merged with the encoded feature map to obtain the intermediate layer's decoded feature map. This intermediate layer's decoded feature map is then input into the feature decoder to obtain the decoded feature map. The corresponding relationship in this process is as follows: ； in, Represents the encoded feature map, This indicates an upsampling operation based on transpose convolution. Indicates step size is Convolution kernel is transposed convolution, This indicates that the convolution has been performed using a transposed convolution with a kernel size of 2 and a stride of 2. This indicates the number of layers in the decoder.
6. The infrared small target detection method based on adaptive partial self-attention according to claim 5, characterized in that, In step 5, the decoded feature map is input into the prediction head to obtain the latent prediction image. The latent prediction images are then integrated to obtain the final detection image. The corresponding relationship in this process is as follows: ； in, Represents the potential prediction image, This indicates that the process has been performed using the Sigmoid activation function. This indicates that the data has undergone upsampling processing using bilinear interpolation. This represents the final detected image.
7. An infrared small target detection system based on adaptive partial self-attention, characterized in that, The system employs the infrared small target detection method based on adaptive partial self-attention as described in any one of claims 1 to 6 above, and the system comprises: The feature encoding module is used for: A single-frame infrared image is processed sequentially through a dual residual attention block and a feature encoder to obtain an encoded feature map; Max pooling downsampling is performed on the coding feature maps at different levels to obtain coding features. The coding features are then concatenated along the channel dimension to obtain a multi-scale high-level coding feature map. An adaptive partial self-attention module is used for: The multi-scale high-level coding feature map is subjected to channel segmentation to obtain the segmented multi-scale high-level coding feature map. Self-attention is then calculated on the segmented multi-scale high-level coding feature map to obtain the global attention map. The global attention map and the multi-scale high-level coding feature map are concatenated along the channel dimension to obtain a fused feature map. The fused feature map is then processed by convolutional layers and multilayer perceptrons to obtain a global context long-range feature map. The feature decoding module is used for: The global context long-range feature map is upsampled based on transpose convolution and then merged with the encoded feature map to obtain the intermediate layer's decoded feature map. The intermediate layer's decoded feature map is then input into the feature decoder to obtain the decoded feature map. Image detection module, used for: The decoded feature map is input into the prediction head to obtain the latent prediction image. The latent prediction images are then integrated to obtain the final detection image. The target is tracked based on the final detected image.