An infrared small target detection method and system based on Mamba2

Abstract

The present application relates to the technical field of infrared image target detection, in particular to an infrared small target detection method and system based on Mamba2. The method adopts a lightweight encoder-decoder network structure, introduces Mamba2 sequence units and channel attention mechanisms in the encoding stage to realize efficient feature extraction and background interference suppression of small targets in infrared images; in the decoding stage, an adaptive feature propagation module and a parallel decoding structure are constructed to sparsely enhance multi-scale semantic information and strengthen the response expression of target regions combined with a spatial attention mechanism. The method has strong global modeling capability and local perception capability, effectively improving the detection accuracy and robustness of small targets in infrared images. Experimental results show that the present application is superior to existing mainstream methods on various public infrared small target data sets, can realize high-precision detection and real-time inference of small targets under complex background interference, and has good engineering practical value.

Description

Technical Field

[0001] This invention relates to the field of computer vision and target detection technology, and in particular to an infrared small target detection method based on Mamba2. Background Technology

[0002] Infrared small target detection technology is widely used in military early warning, aviation surveillance, satellite remote sensing, and border security. Due to its advantages such as all-weather operation, strong penetration, and high concealment, it has gradually become an important component of intelligent sensing systems. However, small targets in infrared images typically exhibit characteristics such as small size, low signal-to-noise ratio, and weak contrast, and are easily affected by complex backgrounds, posing serious challenges to the accuracy and robustness of traditional detection algorithms.

[0003] Existing infrared small target detection methods can be mainly divided into two categories: models based on traditional image processing and models based on deep learning. Traditional methods, such as those based on filters, salient region extraction, or local contrast enhancement, have advantages such as simplicity and real-time performance. However, they are prone to false detections or missed detections when faced with complex backgrounds such as sea surface waves, ground hotspots, and cloud textures. To address this, in recent years, various detection models that integrate multi-scale features and utilize non-local information have emerged to improve the ability to discriminate targets.

[0004] With the development of convolutional neural networks (CNNs) and attention mechanisms, deep learning-based infrared small target detection methods have made significant progress. These methods can automatically learn the latent spatial structure and semantic information in images, significantly improving detection accuracy. However, CNNs suffer from limited receptive fields, weak local modeling capabilities, and insufficient global context modeling, especially when dealing with small targets, where target information is easily overwhelmed by downsampling operations in deep networks. Furthermore, excessive model parameters and computational complexity limit their deployment on resource-constrained platforms. To address these issues, researchers have introduced novel structures such as Transformer architectures, nonlocal mechanisms, and state-space modeling to capture long-distance dependencies in images and achieve more efficient global feature modeling. However, these models still suffer from insufficient target response, weak background suppression, and low detection efficiency in infrared image scenarios, making it difficult to balance detection accuracy and inference speed. Therefore, there is an urgent need to design an infrared small target detection method with a lightweight structure, strong feature representation capabilities, and high detection accuracy to meet the practical application needs in complex scenarios. Summary of the Invention

[0005] This invention addresses the problems of difficult target feature extraction, high false alarm rate, and high model complexity in existing infrared small target detection methods. It proposes an infrared small target detection method and system based on Mamba2 state space modeling to solve these technical problems. The invention aims to integrate the advantages of sequential state space models and convolutional neural networks to achieve accurate detection and background suppression of small targets, improving detection accuracy and robustness while ensuring real-time performance.

[0006] To achieve the above objectives, the technical solution is as follows: This invention provides an infrared small target detection method, comprising the following steps:

[0007] Step 1: Preprocess the input single-frame infrared image and embed it with patches to generate a standard input image;

[0008] Step 2: Input the standard image into the dual-branch Mamba encoder to extract multi-scale features layer by layer. The encoder includes multiple levels of Mamba encoding modules that combine channel attention.

[0009] Step 3: Input the high-level features output by the encoder into the parallel decoder for step-by-step upsampling to reconstruct the spatial resolution;

[0010] Step 4: Input the fused feature map output by the decoder into the spatial attention enhancement module to strengthen the feature representation of the target region;

[0011] Step 5: Input the enhanced features into the mask prediction module to generate a target confidence map or segmentation mask map with the same size as the input image, which is used as the infrared small target detection result.

[0012] Furthermore, step 1 is implemented through a patch encoding module, which includes two parts: a local vector generation branch and a global vector generation branch. The local vector generation branch includes a convolutional layer, a normalization layer, a ReLU activation function, and an Unfold operation to obtain a local feature vector. The global vector generation branch has multiple convolutional modules, each of which includes a convolutional layer, a normalization layer, and a ReLU activation function to obtain a global feature vector.

[0013] Furthermore, each layer of the dual-branch Mamba encoder employs a local-global dual-branch structure to achieve efficient extraction and fusion of weak target features. The input features are divided into two paths: the local branch divides the input features into several local blocks of a fixed size, and each local block is independently input into the isomorphic CAMamba module for local feature modeling, extracting fine-grained local semantic information; the global branch directly sends the complete input features into the CAMamba module for global state modeling, capturing long-range dependencies and macroscopic structural features. After the features from both paths are extracted, they are fused using an element-wise addition method to obtain the enhanced output features.

[0014] Furthermore, the CAMamba module includes a Mamba2 two-dimensional state space modeling unit and a channel attention module. The state update method of the Mamba2 two-dimensional state space modeling unit is as follows:

[0015]

[0016] in, This represents the state value of pixel (i,j) in the k-th direction, including the horizontal, vertical, or diagonal directions. As input features, , , , Update the parameter matrix for the state.

[0017] Furthermore, the multi-scale features extracted by the encoder are further fused by the decoder to restore the output resolution, thereby generating the target segmentation result; the decoder contains two key modules: an adaptive feature propagation module and a multilayer perceptron mapping module; let... and It is the output of the adjacent layer of the dual-branch Mamba encoder. It is a deeper level of output. This is the output of a shallower layer. The adaptive feature propagation module executes between two adjacent layers of the dual-branch Mamba encoder, first receiving features from a pair of adjacent network depths. and ,like If it does not exist, i.e., the initial layer, then directly... As the basic input; otherwise, for Interpolation operations and Convolution makes it with Aligning the data in terms of spatial dimensions and stitching them together, then applying convolution and activation functions, yields a saliency map. Next, from Select several positions with the highest foreground response and then... Max pooling is performed to roughly locate response peaks, and then the top-K pixels with the largest values ​​are selected as representative points of the salient target region. Then, high-level features Upsampling and lower layer Alignment, at a selected position, adds information from higher-level features to lower-level features, achieving cross-level semantic injection;

[0018] For the original features output by the dual-branch Mamba encoder, a lightweight multilayer perceptron mapping module is first applied to process them as one type of feature. The feature set after sparse enhancement by the adaptive feature propagation module is used as another type of feature. Then, the two types of features are concatenated and spliced ​​in the channel dimension to form a fused feature.

[0019] Furthermore, the mask prediction module consists of a single... The system consists of convolutional layers, BatchNorm, and ReLU activation, followed by a Dropout layer to prevent overfitting, and finally a... Convolution maps the channel dimension to the number of target categories, which is a binary classification of foreground / background for infrared small target detection. The feature map output by the mask prediction module is then interpolated and upsampled to the same resolution as the input image, serving as the final small target detection result mask.

[0020] Furthermore, the spatial attention enhancement module generates a spatial attention map through the following mechanism to strengthen the feature representation of the target region;

[0021] (1) Perform max pooling and average pooling on the fused feature maps respectively;

[0022] (2) After concatenating the pooling results along the channel dimension, input them into a set of convolutional layers and sigmoid activation functions to obtain the spatial attention map. ;

[0023] (3) Original feature map and The output after multiplication is an enhanced feature map. :

[0024]

[0025] in This indicates element-wise multiplication.

[0026] Furthermore, the encoder comprises four levels of Mamba encoding modules incorporating channel attention, with spatial output dimensions as follows:

[0027]

[0028] Where H and W are the height and width of the input image, respectively.

[0029] Furthermore, it also includes using evaluation metrics such as precision, recall, F1 score, crossover ratio, false alarm rate, and processing frame rate to quantitatively evaluate and analyze the infrared small target detection results.

[0030] The present invention also provides an infrared small target detection system based on Mamba2, comprising: a processor and a memory, wherein the memory is used to store program instructions, and the processor is used to call the stored instructions in the memory to execute an infrared small target detection method based on Mamba2 as described above.

[0031] Compared with existing technologies, the beneficial effects are:

[0032] (1) This invention integrates the global modeling capability based on the state-space model with the local feature extraction capability of the convolutional neural network, which significantly improves the extraction effect of weak target features. In particular, by extending the Mamba2 sequence unit in two dimensions, it realizes long-distance dependency modeling of information in the image plane, which enhances the comprehensive representation capability of target details and global background.

[0033] (2) The introduction of the channel attention mechanism effectively suppressed background redundancy information and enhanced the target-related channels. In complex backgrounds, compared with the model that did not use channel attention, the false alarm rate and false negative rate of infrared small targets were significantly reduced.

[0034] (3) The proposed adaptive feature propagation module can select and strengthen key target regions to address the problem of insufficient response to weak targets in high-level semantic features, thereby achieving sparse enhancement of the target signal. This targeted enhancement method improves the response strength of weak targets in the decoding stage, making the detection results more accurate.

[0035] (4) The overall network adopts a lightweight encoder-decoder architecture, which has a lower computational complexity than the scheme that introduces global self-attention, thus improving the real-time performance of detection while ensuring accuracy. Experimental results show that the method of the present invention has achieved excellent detection performance on a variety of infrared image datasets, with significant improvements in metrics such as Precision, Recall, and F1, enabling more stable and reliable detection of weak targets in infrared images. Attached Figure Description

[0036] Figure 1 The schematic diagram of the overall structure of the infrared weak target detection network provided by the present invention shows the architectural relationship between the dual-branch encoder, parallel decoder and mask prediction module based on Mamba2.

[0037] Figure 2 This is a schematic diagram of the patch encoding module provided by the present invention;

[0038] Figure 3 The schematic diagram of the dual-branch Mamba encoder module structure provided by the present invention includes a convolutional branch, a Mamba2 sequence branch, and a channel attention fusion design.

[0039] Figure 4This is a schematic diagram of the adaptive feature propagation module provided by the present invention, which enhances the response to weak targets through feature selection, upsampling and fusion.

[0040] Figure 5 This is a flowchart of an embodiment of the present invention.

[0041] Figure 6 This is a visualization of typical scenarios in the experiments of this invention. Detailed Implementation

[0042] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings. Obviously, the following embodiments are only used to explain a part of the present invention and do not limit the entire scope of the present invention.

[0043] This invention provides an infrared small target detection method based on Mamba2, the overall architecture of which is as follows: Figure 1 As shown, the method of this invention adopts an encoder-decoder structure, comprising three parts: a dual-branch Mamba encoder, a parallel decoder, and a mask prediction module. The encoder downsamples the input infrared image layer by layer to extract multi-scale features; the decoder upsamples the high-level semantic features output by the encoder step by step and fuses low-level detail information to restore spatial resolution; the mask prediction module processes the fused features output by the decoder and outputs the segmentation mask where the target is located.

[0044] like Figure 5 As shown, the present invention provides an infrared small target detection method based on Mamba2, comprising the following steps:

[0045] Step 1: Preprocess the input single-frame infrared image and embed it with patches to generate a standard input image;

[0046] Step 2: Input the standard image into the dual-branch Mamba encoder to extract multi-scale features layer by layer. The encoder includes multiple levels of Mamba encoding modules that combine channel attention.

[0047] Step 3: Input the high-level features output by the encoder into the parallel decoder for step-by-step upsampling to reconstruct the spatial resolution;

[0048] Step 4: Input the fused feature map output by the decoder into the spatial attention enhancement module to strengthen the feature representation of the target region;

[0049] Step 5: Input the enhanced features into the mask prediction module to generate a target confidence map or segmentation mask map with the same size as the input image, which is used as the infrared small target detection result.

[0050] First, the input infrared image is divided into fixed-size image blocks by a patch encoding module, and initial feature representations are extracted, such as... Figure 2As shown, this module converts the original infrared image into a sequence of visual tokens with initial semantic representations. It consists of two parts: a local vector generation branch and a global vector generation branch. The local vector generation branch includes convolutional layers, normalization layers, ReLU activation functions, and an unfold operation, ultimately obtaining local feature vectors. The global vector generation branch has multiple convolutional modules, each including a convolutional layer, a normalization layer, and a ReLU activation function, ultimately obtaining global feature vectors. These two branches aim to construct basic image units and high-level abstract semantics, respectively, to support the subsequent two-branch Mamba modeling process.

[0051] The encoder incorporates channel attention mechanisms at each layer to enhance the local contrast of small targets relative to the background. Simultaneously, it performs regionalization processing on the feature maps during encoding to strengthen the feature representation of the target region. At the decoding end, a dual-path parallel decoding structure is designed to fully utilize multi-scale features and suppress redundant noise information. The multilayer perceptron directly upsamples and recovers the multi-scale features from the encoder, while the auxiliary branch selects and strengthens salient regions in high semantic features through an adaptive feature propagation mechanism before entering the upsampling module. Furthermore, a spatial attention enhancement module is introduced after fusion to further highlight the target location and suppress background interference, thereby improving the detectability and localization accuracy of small targets.

[0052] Dual-branch Mamba encoder design: such as Figure 3 The diagram shows the internal structure of the Mamba encoder, with the upper part illustrating the local-global dual-branch processing flow and the lower part showing the internal details of a single CAMamba module. This invention designs an encoder with a local-global dual-branch fusion structure, fully combining the advantages of fine-grained local perception and global semantic modeling, while introducing a channel attention mechanism to further enhance the feature response of small targets. Each Mamba encoding layer employs a local-global dual-branch structure to achieve efficient extraction and fusion of features from weak targets. Input features are divided into two paths: the local branch divides the input features into several local blocks of a fixed size, and each local block is independently input into an isomorphic CAMamba module for local feature modeling, extracting fine-grained local semantic information. The global branch directly feeds the complete input features into the module for global state modeling, capturing long-range dependencies and macroscopic structural features. After the features from both paths are extracted, they are fused using element-wise addition to obtain the enhanced output features. Within each layer, the CAMamba module of this invention consists of two parts in its overall structure: the left side is a global dependency extraction part based on Mamba2 sequence modeling, which processes token sequences as units, and is specifically composed of a state space modeler and an inverse residual feedforward network; the right side is a local feature enhancement part that combines channel attention mechanism, which performs detail extraction and robustness enhancement in the local neighborhood.

[0053] In this embodiment, the CAMamba module includes a Mamba2 two-dimensional state space modeling unit and a channel attention module. The state update method of the Mamba2 two-dimensional state space modeling unit is as follows:

[0054]

[0055] in, This represents the state value of pixel (i,j) in the k-th direction, including the horizontal, vertical, or diagonal directions. As input features, , , , Update the parameter matrix for the state.

[0056] The channel attention mechanism includes: First, performing global average pooling on the input features to obtain the global description vector for each channel. Then, a weight vector with the same number of feature channels as the original is generated through two layers of fully connected neurons. The features are then normalized to the 0-1 range using the Sigmoid function; finally, each channel feature is multiplied by the corresponding... The coefficients enhance the target-related channels and suppress the background channels. Channel attention fusion makes the features output by the dual-branch Mamba encoder more focused on the target signal.

[0057] Through the aforementioned two-dimensional recursive model, the state space unit considers information from both horizontal and vertical neighborhoods simultaneously when traversing the image, enabling the model to capture the long-range dependencies between the target and the background in both dimensions. Furthermore, a multi-directional selective scanning mechanism is employed: the state is updated separately along the horizontal, vertical, and two diagonal directions, and the state results from all four directions are fused, thus avoiding the potential for missing long-range information in a single path. The combination of the aforementioned Mamba2 sequence unit and the channel attention mechanism constitutes the CAMamba module, used by the encoder to extract features from weak targets.

[0058] In this embodiment, the encoder contains a total of 4 Mamba coding layer structures, and their spatial output dimensions are as follows:

[0059]

[0060] Where H and W are the height and width of the input image, respectively.

[0061] Parallel Decoder Design: This embodiment illustrates the structure of the decoder and the implementation process of the adaptive feature propagation module with reference to the accompanying drawings. Figure 1As shown, the multi-scale features extracted by the encoder need to be further fused by the decoder to restore the output resolution, thereby generating the target segmentation result. This embodiment designs a parallel decoding architecture comprising two key modules: an adaptive feature propagation module and a multilayer perceptron mapping module. The adaptive feature propagation module executes between two adjacent layers of the encoder, as shown... Figure 4 As shown, the adaptive feature propagation module first receives a pair of features from adjacent network depths, if... If it does not exist, i.e., the initial layer, then directly... As the basic input; otherwise, for Interpolation operations and Convolution makes it with Aligning the data in terms of spatial dimensions and stitching them together, then applying convolution and activation functions, yields a saliency map. Next, from Select several positions with the highest foreground response and then... Max pooling is performed to roughly locate response peaks, and then the top-K pixels with the largest values ​​are selected as representative points of the salient target region. Then, high-level features Upsampling and lower layer Alignment is achieved by adding information from higher-level features to lower-level features at these selected locations, thus enabling cross-level semantic injection.

[0062] The decoder's fused output is processed by a spatial attention enhancement module: After processing by the adaptive feature propagation module, the decoder simultaneously obtains two sets of multi-scale features: one is the original feature set directly output by the encoder, and the other is the feature set after sparse enhancement by the adaptive feature propagation module. For the original features output by the encoder, a lightweight multilayer perceptron mapping module is first applied. Then, the two types of features are concatenated and concatenated along the channel dimension to form fused features, which are then processed by a spatial attention module. Finally, a lightweight fully convolutional prediction head is used to generate the final target segmentation mask. This prediction head (corresponding to...) Figure 1 The fully connected layer in the middle is very concise, consisting of only one layer. The system consists of convolutional layers, BatchNorm, and ReLU activation, followed by a Dropout layer to prevent overfitting, and finally a... Convolution maps the channel dimension to the number of target categories, resulting in foreground / background binary classification for infrared small target detection. The feature map output by the prediction head is then interpolated and upsampled to the same resolution as the input image, serving as a mask for the final small target detection result.

[0063] The spatial attention enhancement module generates spatial attention graphs through the following mechanism:

[0064] (1) Perform max pooling and average pooling on the fused feature map;

[0065] (2) After concatenating the pooling results along the channel dimension, input them into a set of convolutional layers and sigmoid activation functions to obtain the spatial attention map. ;

[0066] (3) Original feature map and The output after multiplication is an enhanced feature map. :

[0067]

[0068] in This indicates element-wise multiplication.

[0069] We ultimately chose to use U-Net, FCN, ISNet, AGPCNet, and MIM-ISTD methods on a ground / airborne small aircraft dataset to validate the state-of-the-art and effectiveness of our proposed method. The visualization results in typical scenarios are shown in the following figures. Figure 6 As shown in the table below, this invention employs several commonly used evaluation metrics in this field, including precision, recall, F1 score, intersection-over-union ratio (IoU), false alarm rate (Fa), and frame rate (FPS), for quantitative evaluation and analysis, along with several other infrared small target detection methods. The results are shown in the table below:

[0070]

[0071] As shown in the table, the proposed model exhibits significant advantages across several key metrics. Its Recall score of 0.895 is the highest among all compared methods, demonstrating its exceptional target coverage capability. The False Alarm Rate (FA) of only 0.018 is also the best in the table, fully demonstrating the proposed method's ability to suppress false detections in complex contexts. The F1-score of 0.832 surpasses mainstream methods such as AGPCNet, placing it in a leading position. The F1-score reflects the model's ability to balance accuracy and recall, further validating its ability to recover boundaries and optimize response quality for weak targets. This is primarily due to the adaptive feature propagation module introduced in the decoder, which significantly improves the clarity and localization consistency of the target response region.

[0072] U-Net and FCN, as traditional semantic segmentation models with shallow structures, struggle to extract fine-grained target features and perform poorly in separating weak targets from the background, leading to significant false detections. ISNet and AGPCNet, dedicated methods for infrared small target detection, have achieved performance improvements through multi-scale modeling and attention mechanisms. AGPCNet excels in the Fa metric, sacrificing some recall for false alarm suppression; ISNet achieves a recall of 0.872, but lacks deep structure optimization, resulting in insufficient adaptability to complex backgrounds and mediocre precision and IoU. MiM-ISTD uses a Mamba structure to model long-range dependencies, achieving a recall of 0.887, demonstrating strong overall target coverage. However, due to the absence of the proposed channel attention and sparse feature propagation structure, it suffers from weaknesses in background suppression, and detection results are still susceptible to noise interference.

[0073] On the other hand, embodiments of the present invention also provide an infrared small target detection system based on Mamba2, comprising: a processor and a memory, wherein the memory is used to store program instructions, and the processor is used to call the stored instructions in the memory to execute an infrared small target detection method based on Mamba2 as described above.

[0074] The specific embodiments described herein are merely illustrative examples of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

Claims

1. A method for detecting small infrared targets based on Mamba2, characterized in that, Includes the following steps: Step 1: Preprocess the input single-frame infrared image and embed it with patches to generate a standard input image; Step 2: Input the standard image into the dual-branch Mamba encoder to extract multi-scale features layer by layer. The encoder includes multiple levels of Mamba encoding modules that combine channel attention. The dual-branch Mamba encoder employs a local-global dual-branch structure in each layer to achieve efficient extraction and fusion of features from weak targets; The input features are divided into two paths: the local branch divides the input features into several local blocks of a fixed size, and each local block is independently input into the isomorphic CAMamba module for local feature modeling to extract fine-grained local semantic information; the global branch directly feeds the complete input features into the CAMamba module for global state modeling to capture long-range dependencies and macro-structural features; after the features from both paths are extracted, they are fused using an element-wise addition method to obtain the enhanced output features. The CAMamba module includes the Mamba2 two-dimensional state space modeling unit and the channel attention module. The state update method of the Mamba2 two-dimensional state space modeling unit is as follows: in, This represents the state value of pixel (i,j) in the k-th direction, including the horizontal, vertical, or diagonal directions. As input features, , , , Update the parameter matrix for the state; Step 3: Input the high-level features output by the encoder into the parallel decoder for step-by-step upsampling to reconstruct the spatial resolution; The multi-scale features extracted by the encoder are further fused by the decoder to restore the output resolution, thereby generating the target segmentation result; the decoder contains two key modules: an adaptive feature propagation module and a multilayer perceptron mapping module; let... and It is the output of the adjacent layer of the dual-branch Mamba encoder. It is a deeper level of output. This is the output of a shallower layer. The adaptive feature propagation module executes between two adjacent layers of the dual-branch Mamba encoder, first receiving features from a pair of adjacent network depths. and ,like If it does not exist, i.e., the initial layer, then directly... As the basic input; otherwise, for Interpolation operations and Convolution makes it with Aligning the data in terms of spatial dimensions and stitching them together, then applying convolution and activation functions, yields a saliency map. Next, from Select several positions with the highest foreground response and then... Max pooling is performed to roughly locate response peaks, and then the top-K pixels with the largest values ​​are selected as representative points of the salient target region. Then, high-level features Upsampling and lower layer Alignment, at a selected position, adds information from higher-level features to lower-level features, achieving cross-level semantic injection; For the original features output by the dual-branch Mamba encoder, a lightweight multilayer perceptron mapping module is first applied to process them as one type of feature. The feature set after sparse enhancement by the adaptive feature propagation module is used as another type of feature. Then, the two types of features are concatenated and spliced ​​in the channel dimension to form a fused feature. Step 4: Input the fused feature map output by the decoder into the spatial attention enhancement module to strengthen the feature representation of the target region; Step 5: Input the enhanced features into the mask prediction module to generate a target confidence map or segmentation mask map with the same size as the input image, which is used as the infrared small target detection result.

2. The infrared small target detection method based on Mamba2 as described in claim 1, characterized in that: Step 1 is implemented through a patch encoding module, which includes two parts: a local vector generation branch and a global vector generation branch. The local vector generation branch includes convolutional layers, normalization layers, ReLU activation functions, and an unfold operation to obtain local feature vectors. The global vector generation branch has multiple convolutional modules, each including a convolutional layer, a normalization layer, and a ReLU activation function to obtain global feature vectors.

3. The infrared small target detection method based on Mamba2 as described in claim 1, characterized in that: The mask prediction module consists of one The system consists of convolutional layers, BatchNorm, and ReLU activation, followed by a Dropout layer to prevent overfitting, and finally a... Convolution maps the channel dimension to the number of target categories, which is a binary classification of foreground / background for infrared small target detection. The feature map output by the mask prediction module is then interpolated and upsampled to the same resolution as the input image, serving as the final small target detection result mask.

4. The infrared small target detection method based on Mamba2 as described in claim 1, characterized in that: The spatial attention enhancement module generates a spatial attention map through the following mechanism to enhance the feature representation of the target region; (1) Perform max pooling and average pooling on the fused feature maps respectively; (2) After concatenating the pooling results along the channel dimension, input them into a set of convolutional layers and sigmoid activation functions to obtain the spatial attention map. ; (3) Original feature map and The output after multiplication is an enhanced feature map. : in This indicates element-wise multiplication.

5. The infrared small target detection method based on Mamba2 as described in claim 1, characterized in that: The encoder contains four levels of Mamba encoding modules that incorporate channel attention, with spatial output dimensions as follows: Where H and W are the height and width of the input image, respectively.

6. The infrared small target detection method based on Mamba2 as described in claim 1, characterized in that: It also includes using evaluation metrics such as precision, recall, F1 score, crossover ratio, false alarm rate, and processing frame rate to quantitatively evaluate and analyze the infrared small target detection results.

7. An infrared small target detection system based on Mamba2, characterized in that, include: The processor and memory, wherein the memory is used to store program instructions, and the processor is used to call the stored instructions in the memory to execute the infrared small target detection method based on any one of claims 1-6.

Images