Infrared small target detection method and system

Abstract

The present application relates to the technical field of computer vision, and more particularly to an infrared small target detection method and system. In view of the problem of single information dimension of infrared single-channel image, local variance and luminance gradient feature channels are constructed from original luminance features, and adaptive fusion is completed to fully exploit the multi-dimensional exclusive semantic features of infrared images. In the feature extraction stage, an exclusive branch is constructed to capture the fine-grained spatial detail features of small infrared targets smaller than 4x4 pixels, and a three-fusion gate mechanism combining input features, infrared segmentation position encoding and historical state is introduced to capture the long-distance association between global background and infrared small targets. Based on the luminance features of infrared images, a luminance mask that fits the infrared scene is generated to achieve accurate semantic filtering and effectively distinguish targets from similar gray backgrounds. Meanwhile, a channel-spatial joint attention mechanism is introduced in the two key stages of feature extraction and cross-scale fusion to effectively improve the accuracy of infrared small target detection.

Description

Technical Field

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

[0002] Infrared single-channel images are widely used in security monitoring and other fields for small target detection. These images possess unique grayscale and brightness information, enabling them to capture target contours and radiation characteristics in complex lighting conditions such as low light and nighttime. This provides core data support for small target identification; the brightness information reflects the radiation difference between the target and the background, while the spatial information conveys the target's position and morphological features.

[0003] Methods for infrared small target detection using single-channel infrared images can generally be divided into those based on traditional image processing algorithms and those based on simple neural networks. Traditional image processing-based methods primarily rely on classical digital image processing theory, utilizing the brightness and grayscale information and pixel spatial location information in the single-channel infrared image to construct feature extraction and target segmentation modules, thereby effectively distinguishing targets from the background in the infrared image. Commonly used image processing methods include thresholding, morphological operations, and gradient filtering. Other image processing methods, such as spatial domain filtering algorithms, frequency domain enhancement algorithms, and edge detection algorithms, also play important roles. These methods typically extract the brightness difference features and spatial morphological features of targets in the infrared image, fusing these different features to ensure the identifiability of small targets.

[0004] With the continuous advancement of deep learning theory, several deep network models have been applied to the research of infrared small target detection, such as Convolutional Neural Networks (CNN), Selective State Space Models (Mamba), and Long Short-Term Memory (LSTM). These methods construct complex deep network structures to perform deep feature mining on single-channel infrared images, comprehensively representing and describing the core discriminative features of infrared small targets from multiple dimensions. Therefore, an increasing number of deep learning-based infrared small target detection methods have been proposed.

[0005] However, existing infrared small target detection methods based on deep network models mostly employ conventional fusion methods such as direct cross-scale stitching, simple weighted summation, or unidirectional disordered upsampling in their feature fusion units. This results in a disconnect between shallow high-resolution detail features and deep high-dimensional semantic features. Features at different resolution levels cannot be effectively adapted and coupled, making it difficult to form synergistic and complementary feature representations. Weak target features are easily masked by complex background clutter, leading to extremely low overall feature utilization. Furthermore, these methods have very low distinguishability between targets and similar grayscale backgrounds, failing to accurately remove background interference noise. This easily leads to missed detections of small targets and misidentification of background clutter as targets, ultimately resulting in detection accuracy and robustness that fail to meet the application requirements of real-world scenarios. Summary of the Invention

[0006] Therefore, the technical problem to be solved by the present invention is to overcome the defects of existing infrared small target detection methods based on deep network models, which have poor accuracy in infrared small target detection because the shallow high-resolution detail features and deep high-dimensional semantic features are separated in the multi-scale feature fusion mechanism, the features at different resolution levels cannot be effectively adapted and coupled, and the background interference noise cannot be accurately removed.

[0007] To address the aforementioned technical problems, this invention provides an infrared small target detection method, comprising: The preprocessed infrared single-channel grayscale image is used as the input image of the feature extraction unit, and features at different scales are extracted using the feature extraction unit. Features at different scales are processed by a feature fusion unit to obtain a fused and enhanced feature map, including: Each scale feature is extracted by using depthwise separable convolution and 1×1 convolution; Based on the local brightness mean and local brightness variance of each scale feature, a brightness mask for each scale is generated. After multiplying the spatial semantic features of each scale element-wise with the brightness mask, the binary semantic map of each scale is obtained by passing it through the Sigmoid activation function. Based on the binary semantic graph at each scale, obtain the global semantic confidence weight for each scale; Each scale feature is multiplied element-wise by the binary semantic graph of that scale, and then multiplied element-wise by the global semantic confidence weight of that scale to obtain the weighted feature of each scale. After upsampling the weighted features of each scale to the maximum resolution scale among the features of each scale, the channel-spatial attention module is used to obtain the channel-spatial attention features of each scale. Based on the binary semantic map and channel-space attention features at each scale, a spatial attention map at each scale is generated. The spatial attention maps at each scale are then spliced ​​along the channel dimension to obtain a multi-scale fusion feature map. The multi-scale fused feature map is passed through the improved three-fusion gate Mamba module to obtain the fused enhanced feature map; The fused and enhanced feature maps are then processed through an eight-domain clustering module to obtain infrared small target detection results.

[0008] Preferably, the process of acquiring the preprocessed infrared single-channel grayscale image includes: The infrared single-channel grayscale image is normalized to obtain the normalized infrared single-channel grayscale image. A 3×3 Gaussian filter was used to denoise the normalized infrared single-channel grayscale image to obtain a preprocessed infrared single-channel grayscale image.

[0009] Preferably, the method for optimizing the preprocessed infrared single-channel grayscale image to obtain the enhanced input image for the feature extraction unit includes: The preprocessed infrared single-channel grayscale image is used as the original brightness feature; A 3×3 sliding window is used to traverse all pixels of the preprocessed infrared single-channel grayscale image, and the local brightness variance is calculated to obtain the local variance features. Sobel gradient calculation is performed on the preprocessed infrared single-channel grayscale image to calculate its gradient in the horizontal and vertical directions respectively. The brightness gradient features are obtained by fusing the horizontal and vertical gradients of the preprocessed infrared single-channel grayscale image. The original brightness features, local variance features, and brightness gradient features are concatenated along the channel dimension to obtain the enhanced input image of the infrared small target detection model.

[0010] Preferably, the feature extraction unit consists of multiple parallel branches, each of which is a densely nested attention network that extracts features at different scales; among them, the shallow branch with the highest resolution has a size of 1 / 2 of the input image and 16 channels.

[0011] Preferably, the densely nested attention network includes: multiple layers connected in sequence, each layer including a convolutional layer and a channel-space attention module connected in sequence, and dense skip connections and bidirectional interactive skip connections are used between the layers.

[0012] Preferably, the last convolutional layer in the deep branch with the lowest resolution is replaced with an improved three-fusion gate Mamba module to obtain the target feature extraction unit.

[0013] Preferably, the method for obtaining the output feature map of the improved three-gate Mamba module by passing the input feature map through the improved three-gate Mamba module includes: The input feature map is flattened in the spatial dimension to obtain a sequence of input feature maps; infrared segmentation position encoding is applied to the sequence of multi-scale fusion features to obtain position codes; For each sequence position in the input feature map in sequence form, perform a linear transformation on the input feature value, position code, and historical state information of the previous sequence position to obtain the linearly transformed input feature value, position code, and historical state information of the previous sequence position. The updated input feature values ​​at the current sequence position are passed sequentially through a multilayer perceptron and a softmax activation function to obtain the adaptive fusion weights at the current sequence position. Based on the adaptive fusion weights of the current sequence position, the input feature value position encoding after linear transformation of the current sequence position and the historical state information of the previous sequence position are weighted and summed, and then the forgetting gate vector of the current sequence position is obtained by passing the Sigmoid activation function. Based on the state decay coefficient, the forgetting gate vector of the current sequence position, the input feature value after linear transformation of the current sequence position, and the historical state information after linear transformation of the previous sequence position, the historical state information of the current sequence position is obtained. Based on the feature value after linear transformation of the current sequence position and historical state information, obtain the final feature value of the current sequence position; After all positions in the entire sequence have been processed, the final feature values ​​output from each sequence position are concatenated in the original sequence order and reconstructed into a two-dimensional feature form, resulting in the output feature map of the improved three-way Mamba module.

[0014] Preferably, the formula for obtaining the historical state information of the current sequence position based on the state decay coefficient, the forgetting gating vector of the current sequence position, the input feature value after linear transformation of the current sequence position, and the historical state information after linear transformation of the previous sequence position is as follows: , in, Current sequence position Historical status information Position of the previous sequence Historical state information after linear transformation The state decay coefficient, Current sequence position The forgetting gate vector, The input feature value is the linear transformation of the current sequence position.

[0015] Preferably, the method for obtaining the global semantic confidence weight for each scale based on the binary semantic graph at each scale includes: For each scale of the binary semantic map, pixels with values ​​greater than a set threshold are taken as target pixels, and the ratio of the number of target pixels to the total number of pixels in the binary semantic map is used as the global semantic confidence weight for the corresponding scale.

[0016] The present invention also provides an infrared small target detection system, comprising: The feature extraction module is used to take the preprocessed infrared single-channel grayscale image as the input image of the feature extraction unit, and use the feature extraction unit to extract features at different scales. The feature fusion module is used to pass features of different scales through the feature fusion unit to obtain a fused and enhanced feature map, including: Each scale feature is extracted by using depthwise separable convolution and 1×1 convolution; Based on the local brightness mean and local brightness variance of each scale feature, a brightness mask for each scale is generated. After multiplying the spatial semantic features of each scale element-wise with the brightness mask, the binary semantic map of each scale is obtained by passing it through the Sigmoid activation function. Based on the binary semantic graph at each scale, obtain the global semantic confidence weight for each scale; Each scale feature is multiplied element-wise by the binary semantic graph of that scale, and then multiplied element-wise by the global semantic confidence weight of that scale to obtain the weighted feature of each scale. After upsampling the weighted features of each scale to the maximum resolution scale among the features of each scale, the channel-spatial attention module is used to obtain the channel-spatial attention features of each scale. Based on the binary semantic map and channel-space attention features at each scale, a spatial attention map at each scale is generated. The spatial attention maps at each scale are then spliced ​​along the channel dimension to obtain a multi-scale fusion feature map. The multi-scale fused feature map is passed through the improved three-fusion gate Mamba module to obtain the fused enhanced feature map; The detection module is used to obtain infrared small target detection results by passing the fused and enhanced feature map through the eight-neighbor clustering module.

[0017] Compared with the prior art, the above-described technical solution of the present invention has the following advantages: The infrared small target detection method and system described in this invention sequentially performs spatial semantic extraction, brightness mask generation, and semantic filtering operations for each scale feature. A binary semantic map is generated using local brightness features to remove background clutter interference and enhance weak target features. Then, based on the binary semantic map, the global semantic confidence weight for the corresponding scale is calculated, achieving accurate weighted calibration of single-scale features. Next, the weighted features at each scale are uniformly upsampled to the largest resolution scale among all scale features using bilinear interpolation, compensating for the lack of cross-scale feature alignment logic and achieving accurate spatial dimension adaptation of features at different resolution levels. Subsequently, the aligned features at each scale are... The target feature weights are further enhanced and background interference is suppressed by the channel-spatial attention module to generate channel-spatial attention features at the corresponding scale. Then, the binary semantic map at each scale is combined with the channel-spatial attention features to generate a spatial attention map. The multi-scale spatial attention maps are then stitched together along the channel dimension to break the separation between shallow high-resolution detail features and deep semantic features, achieving efficient coupling and synergistic complementarity between shallow and deep features. Finally, the stitched multi-scale fused feature map is input into the improved three-fusion gate Mamba module to complete subsequent feature enhancement and global correlation modeling, comprehensively improving the utilization rate of multi-scale features and preventing weak target features from being obscured by background clutter. Attached Figure Description

[0018] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein: Figure 1 This is a flowchart illustrating an infrared small target detection method according to the present invention. Detailed Implementation

[0019] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0020] Reference Figure 1 As shown, this embodiment provides an infrared small target detection method, including: Step S1: Acquire an infrared single-channel grayscale image, preprocess the infrared single-channel grayscale image to obtain a preprocessed infrared single-channel grayscale image.

[0021] The size of the infrared single-channel grayscale image is , and It refers to the spatial dimensions of a single-channel infrared grayscale image in each band, namely its height and width. This represents the number of channels in the infrared single-channel grayscale image, which is 1. In this embodiment, the size of the infrared single-channel grayscale image is [size missing]. In infrared small target detection, a small target is defined as a region with a pixel area <100 and an aspect ratio <4, among which "micro-targets <4×4 pixels" are the core detection objects.

[0022] In this embodiment, based on the actual needs of infrared small target detection, the publicly available dataset NUAA-SIRST and the self-developed NUDT-SIRST dataset are selected.

[0023] In this embodiment, the method for preprocessing an infrared single-channel grayscale image specifically includes: Normalize the infrared single-channel grayscale image by normalizing the grayscale values ​​to [-1, 1] to obtain the normalized infrared single-channel grayscale image. The formula is: , in, This represents the original grayscale value of a pixel in an infrared single-channel grayscale image. This represents the minimum grayscale value in a single-channel infrared grayscale image. This represents the maximum grayscale value in a single-channel infrared grayscale image. This refers to the normalized grayscale value of a pixel in an infrared single-channel grayscale image.

[0024] 3×3 Gaussian filter (standard deviation) The normalized infrared single-channel grayscale image is denoised to obtain a preprocessed infrared single-channel grayscale image.

[0025] Step S2: Construct an infrared small target detection model, which includes a feature extraction unit, a feature fusion unit, and an eight-neighborhood clustering module; use the preprocessed infrared single-channel grayscale image as the input image of the feature extraction unit, and use the feature extraction unit to extract features at different scales; Existing infrared small target detection methods often employ simple channel copying or single grayscale value stretching for input preprocessing. Infrared single-channel images have a single spectral dimension, making them susceptible to interference from clouds, noise, and complex background clutter. Small infrared targets smaller than 4×4 pixels have weak features, and it is difficult to accurately distinguish between the target and the background based solely on single brightness information. This results in insufficient basic information for subsequent feature extraction and weak resistance to background clutter interference, thus limiting the improvement of detection accuracy.

[0026] Therefore, in this embodiment, preferably, the method for optimizing the preprocessed infrared single-channel grayscale image to obtain the enhanced input image of the infrared small target detection model includes: The preprocessed infrared single-channel grayscale image is used as the original brightness feature; A 3×3 sliding window is used to traverse all pixels of the preprocessed infrared single-channel grayscale image, and the local brightness variance is calculated to obtain the local variance features. Sobel gradient calculation is performed on the preprocessed infrared single-channel grayscale image to calculate its gradient in the horizontal and vertical directions respectively. The brightness gradient features are obtained by fusing the horizontal and vertical gradients of the preprocessed infrared single-channel grayscale image. The original brightness features, local variance features, and brightness gradient features are concatenated along the channel dimension to obtain the enhanced input image of the infrared small target detection model.

[0027] Specifically, extract the preprocessed infrared single-channel grayscale image. Original brightness characteristics Local variance characteristics Brightness gradient characteristics Then, the three features are concatenated along the channel dimension to obtain a 3-channel enhanced feature map. This completes the initial enhancement of infrared semantics, adapting it to the input requirements of subsequent modules. The formula is: , The size is .

[0028] In this embodiment, the preprocessed infrared single-channel grayscale image is extracted. Original brightness characteristics Local variance characteristics Brightness gradient characteristics The methods include: To preserve the basic brightness information of the image, the preprocessed infrared single-channel grayscale image... As the original brightness feature ,Right now ; The preprocessed infrared single-channel grayscale image is traversed using a 3×3 sliding window. For all pixels, calculate the local luminance variance to obtain the local variance feature. To highlight the compact characteristics of small infrared targets—"high brightness, low variance"—and suppress background clutter areas, the specific method is as follows: Preprocessed infrared single-channel grayscale image Each pixel Take the 3×3 neighborhood centered on this pixel. (Number of pixels in the neighborhood) ), calculate the 3×3 neighborhood centered on this pixel. The local variance is given by the formula: , in, Local variance characteristics medium pixel The value, in pixels 3×3 neighborhood centered The local variance, In pixels 3×3 neighborhood centered The local mean of all pixels within the range. Here are the row coordinates of the pixels. The column coordinates of the pixels. , 3×3 neighborhood The coordinates of each pixel within the range. In pixels 3×3 neighborhood centered The number of pixels within, In pixels 3×3 neighborhood centered Inner The brightness value of each pixel. In pixels 3×3 neighborhood centered Inner row pixels, In pixels 3×3 neighborhood centered Inner Column pixels.

[0029] Preprocessed infrared single-channel grayscale image Perform the Sobel gradient calculation to calculate the gradients in the horizontal and vertical directions, respectively, using the following formula: , , in, Preprocessed infrared single-channel grayscale image medium pixel Horizontal gradient, Preprocessed infrared single-channel grayscale image medium pixel vertical gradient, For the horizontal Sobel operator, , For the vertical Sobel operator, .

[0030] Preprocessed infrared single-channel grayscale image The brightness gradient features are obtained by fusing the horizontal and vertical gradients. This is used to enhance the distinction between the edge contour of small targets and the background. The formula is: , in, Brightness gradient features medium pixel The value of .

[0031] This invention proposes a brightness adaptive channel enhancement method for infrared small target detection. Addressing the issue of limited information dimensions in single-channel infrared images, this method constructs local variance and brightness gradient feature channels from the original brightness features and performs adaptive fusion. It mines multi-dimensional semantic information from the original brightness, local variance, and brightness gradient in single-channel infrared images, fully exploring the multi-dimensional exclusive semantic features of infrared images. This overcomes the limitations of traditional single-channel preprocessing, which relies solely on grayscale values, achieving quality improvement at the feature input level. This lays a high-quality foundation for subsequent feature extraction and enhances the algorithm's ability to resist background clutter interference.

[0032] The shallow branch resolution of existing multi-scale feature extraction architectures is insufficient (most of them start from 1 / 4 of the image size), and there is a lack of fine-grained feature capture branches specifically for tiny infrared targets smaller than 4×4 pixels, which leads to the loss of detailed features of tiny infrared targets and a high false negative rate.

[0033] Therefore, in this embodiment, preferably, the feature extraction unit consists of multiple parallel branches, each of which is a densely nested attention network (DNA-Net) that extracts features at different scales; wherein, the shallow branch with the highest resolution has a size of 1 / 2 of the input image and 16 channels.

[0034] The densely nested attention network consists of multiple layers connected in sequence. Each layer includes a convolutional layer and a channel-spatial attention module (CSAM) connected in sequence. Dense plain skip connections and bi-directional interactive skip connections are used between the layers to achieve cross-scale information interaction and ensure the effective fusion of shallow detail features and deep semantic features. The channel-spatial attention module (CSAM) is used to further enhance the weight of target features, suppress background clutter interference, and specifically adapt to the extraction requirements of weak signals from small infrared targets.

[0035] In this embodiment, the channel attention weights in the channel-spatial attention module are calculated using the following formula: , in, The input features are those of the channel-space attention module. for Channel attention weights, This is a global average pooling operation. It is a multilayer perceptron. This is the Sigmoid activation function.

[0036] The spatial attention weights in the channel-spatial attention module are calculated using the following formula: , in, for Spatial attention weights, It is a 1×1 convolutional layer. This is a global max pooling operation.

[0037] Based on the input features of the channel-spatial attention module, along with its channel attention weights and spatial attention weights, the output features of the channel-spatial attention module are calculated using the following formula: , in, The output features of the channel-spatial attention module, This is for element-wise multiplication.

[0038] In this embodiment, specifically, the number of parallel branches is 6, and the shallowest branch with the highest resolution is the S6 branch, whose size is half of the input image and has 16 channels. It uses high resolution and low channel count, specifically designed to capture tiny infrared target details smaller than 4×4 pixels, in order to adapt to the enhanced edge and variance features after input preprocessing. The remaining branches are densely nested attention networks, which perform downsampling sequentially, with the number of channels increasing progressively and the downsampling factor increasing. Corresponding size arrive The number of channels changes from 32 to 64 to 128 to 256 to 512, satisfying the following formula: , in, For the first The height of each parallel branch The height of the input image. For parallel branch indexes, For the first The width of each parallel branch, The width of the input image. For the first The number of channels in each parallel branch. At that time, the shallowest branch with the highest resolution (S6 branch) is selected. At that time, it corresponds to the deepest branch.

[0039] Since the S6 branch itself serves as the baseline size for upsampling in all subsequent branches, no additional upsampling operation is required. It directly participates in subsequent dynamic semantic weighting and cross-scale fusion, preserving the data throughout the process. High resolution and 16-channel low-dimensional features Existing methods (including CNN and traditional RNN / Transformer) have shortcomings in long-distance correlation modeling: CNN has difficulty capturing the long-distance correlation between the global background and small targets, Transformer has high computational complexity and poor real-time performance, and lacks position encoding and gating mechanisms specific to infrared scenes, making it easy to lose small target features and be affected by clutter interference; to this end, this invention incorporates an improved three-fusion gate Mamba module into the feature extraction unit and feature fusion module.

[0040] To address the shortcomings of existing convolutional structures in global long-distance correlation modeling, this invention replaces the last convolutional layer in the deepest gray branch with the lowest resolution with an improved three-fusion gate Mamba module, while leaving the other branches unchanged, thus obtaining the target feature extraction unit.

[0041] In this embodiment, specifically, the method for obtaining the output feature map of the improved three-gate Mamba module by passing the input feature map through the improved three-gate Mamba module includes: The input feature map is flattened in the spatial dimension to obtain a sequence of input feature maps; infrared segmentation position encoding is applied to the sequence of multi-scale fusion features to obtain position codes; For each sequence position in the input feature map in sequence form, perform a linear transformation on the input feature value, position code, and historical state information of the previous sequence position to obtain the linearly transformed input feature value, position code, and historical state information of the previous sequence position. The updated input feature values ​​at the current sequence position are passed sequentially through a multilayer perceptron and a softmax activation function to obtain the adaptive fusion weights at the current sequence position. Based on the adaptive fusion weights of the current sequence position, the input feature value position encoding after linear transformation of the current sequence position and the historical state information of the previous sequence position are weighted and summed, and then the forgetting gate vector of the current sequence position is obtained by passing the Sigmoid activation function. Based on the state decay coefficient, the forgetting gate vector of the current sequence position, the input feature value after linear transformation of the current sequence position, and the historical state information after linear transformation of the previous sequence position, the historical state information of the current sequence position is obtained. The final feature value of the current sequence position is obtained by weighted fusion based on the forgetting gate vector of the current sequence position, the feature value after linear transformation, and the updated historical state information. After all positions in the entire sequence have been processed, the final feature values ​​output from each sequence position are concatenated in the original sequence order and reconstructed into a two-dimensional feature form, resulting in the output feature map of the improved three-way Mamba module.

[0042] In this embodiment, specifically, in the feature extraction unit, the improved three-way fusion gate Mamba module receives the feature map after processing by the penultimate level channel-spatial attention feature. After converting it into sequence features, the system uses a three-fusion gating mechanism of "input features + infrared segmented position coding + historical state" to accurately distinguish between global background clutter and small target semantic regions, thus avoiding the loss of early small target information.

[0043] The feature map after the penultimate level channel - spatial attention feature processing ( Convert the deepest branch to sequence form (Spatial dimension flattened to sequence length) Channel is the feature dimension ), Perform linear transformations on each branch of the three-way fusion gate: , , , in, Current sequence position The input feature values ​​after linear transformation Current sequence position The input feature values, The characteristic linear transformation matrix, For feature bias terms, Position of the previous sequence Historical state information after linear transformation This refers to the historical state information of the previous sequence position. This is a bias term for historical state information. The linear transformation matrix represents the historical state information. Current sequence position Position encoding after linear transformation Current sequence position Location encoding, This is the positional encoding bias term.

[0044] The updated input feature values ​​at the current sequence position are sequentially passed through a multilayer perceptron and a softmax activation function to obtain the adaptive fusion weights for the current sequence position, as shown in the formula: , in, Current sequence position Adaptive fusion weights, For the Softmax function, It is a multilayer perceptron.

[0045] Based on the adaptive fusion weights of the current sequence position, the input feature value position encoding after linear transformation of the current sequence position and the historical state information of the previous sequence position are weighted and summed, and then the forgetting gate vector of the current sequence position is obtained by passing it through the Sigmoid activation function. The formula is as follows: , in, Current sequence position The forgetting gate vector, It is the Sigmoid activation function. Current sequence position In the adaptive fusion weights, the input features after linear transformation correspond to the weights. Current sequence position The positional encoding corresponding to the linearly transformed weights in the adaptive fusion weights Current sequence position The adaptive fusion weights correspond to the historical state information after linear transformation.

[0046] Based on the state decay coefficient, the forgetting gating vector of the current sequence position, the input feature value after linear transformation of the current sequence position, and the historical state information after linear transformation of the previous sequence position, the historical state information of the current sequence position is obtained, as follows: , in, Current sequence position Historical status information Position of the previous sequence Historical state information after linear transformation The state decay coefficient, , For learnable parameters, Current sequence position The forgetting gate vector, The input feature value is the linear transformation of the current sequence position.

[0047] Based on the feature value after linear transformation of the current sequence position and historical state information, the final feature value of the current sequence position is obtained using the following formula: , in, Current sequence position The final eigenvalues, For the state dimension, These are learnable bias parameters.

[0048] After all positions in the entire sequence have been processed, the final feature values ​​output from each sequence position are concatenated in the original sequence order to reconstruct a two-dimensional feature form. This reconstructed feature form is then performed in the feature extraction unit. The improved three-way Mamba module output feature map is obtained and connected to subsequent jump connection and attention modules, maintaining the integrity of the original network's dense nested structure.

[0049] This invention proposes a three-gating Mamba long-distance correlation modeling method adapted to infrared scenes. It designs a dedicated gating mechanism of "input features + infrared segmented position encoding + historical state" to replace the traditional network convolution module. This method can efficiently capture the long-distance correlation between the global background and infrared small targets, and rely on the Mamba architecture to ensure the real-time performance of model inference. This achieves a balance between high-precision detection and high real-time performance, and accurately suppresses the interference of background clutter on the features of small targets.

[0050] Step S3: The features at different scales are processed by the feature fusion unit to obtain the fused enhanced feature map. The feature fusion unit includes an infrared-specific lightweight semantic branch (PLSS), which includes a parallel convolutional branch and a brightness-aware sub-branch. The preprocessed brightness and variance features are input to form a correspondence, further enhancing the infrared semantic distinction.

[0051] Step S31: Extract the spatial semantic features for each scale by performing depth-separable convolutions and 1×1 convolutions on each scale feature, using the following formula: ,in, For the first Each parallel branch corresponds to a spatial semantic feature of the scale. For 1×1 convolution, It uses a 3×3 depthwise separable convolution to control computational cost while preserving target spatial location information; Step S32: The brightness perception sub-branch generates a brightness mask for each scale based on the local brightness mean and local brightness variance of each scale feature (which is logically consistent with the variance feature of the input preprocessing, forming a full-link brightness feature linkage). The mask value for the target region (high mean, low variance) is approximately 1, and the mask value for the background clutter region (high mean, high variance or low mean, low variance) is approximately 0.1.

[0052] The formula for calculating the mean local brightness of each scale feature is as follows: , The formula for calculating the local brightness variance of each scale feature is as follows: , in, For the first The local brightness mean of the scale feature corresponding to each parallel branch. For the first The local brightness variance of the scale feature corresponding to each parallel branch. 3×3 neighborhood The coordinates of each pixel within the range. Here are the row coordinates of the pixels. The column coordinates of the pixels. For the first Each parallel branch corresponds to a scale feature in pixels. 3×3 neighborhood centered Inner The brightness value of each pixel. In pixels 3×3 neighborhood centered Inner row pixels, In pixels 3×3 neighborhood centered Inner Column pixels.

[0053] The formula for generating a luminance mask for each scale, based on the mean and variance of local luminance for each scale feature, is as follows: ,

[0054] in, For the first The position in the luminance mask corresponding to each parallel branch at each scale The value, For preset adjustment parameters, This is the tanh activation function.

[0055] Step S33: Multiply the spatial semantic features of each scale element-wise with the brightness mask, retaining only the regions whose brightness features match the characteristics of small infrared targets, and then pass the Sigmoid activation function to obtain the binary semantic map for each scale. Thus, the preliminary background filtering formula at the semantic level is: , , in, For the first The binary semantic graph corresponding to each parallel branch scale. , For the first Each parallel branch corresponds to a brightness mask at a specific scale. For the first Each parallel branch corresponds to a spatial semantic feature of the scale. For Hadama accumulation.

[0056] Step S34: Based on the binary semantic map at each scale, obtain the global semantic confidence weight for each scale, including: For each scale of the binary semantic map, pixels with values ​​greater than a set threshold (e.g., 0.5) are designated as target pixels. The ratio of the number of target pixels to the total number of pixels in the binary semantic map is used as the global semantic confidence weight for the corresponding scale. A larger ratio results in a larger weight, ensuring that the small target information in shallow high-resolution features (e.g., S6) is not diluted. The formula is as follows: , in, For the first The global semantic confidence weights for each parallel branch at each scale.

[0057] Step S35: Multiply each scale feature element-wise with the binary semantic map of that scale, and then multiply each feature element-wise with the global semantic confidence weight of that scale to obtain the weighted feature for each scale, achieving local precision enhancement and global scale adaptation. The formula is as follows: , in, For the first Each parallel branch corresponds to a weighted feature of the scale. No. Each parallel branch corresponds to a scale feature.

[0058] Step S36: After upsampling the weighted features of each scale to the maximum resolution scale among the features of each scale, the channel-spatial attention module is used to obtain the channel-spatial attention features of each scale. In this embodiment, the weighted features of each scale are... ( Unified upsampling to the high-resolution reference (size) of the S6 branch In this embodiment, bilinear interpolation is used to achieve upsampling, ensuring that the spatial dimensions of all scale features are consistent, laying the foundation for subsequent attention enhancement and channel stitching, while preserving pixel-level details of small targets.

[0059] The features after upsampling are denoted as ,satisfy: , For the first The weighted features after upsampling at the scale corresponding to each parallel branch; Scale features after all upsampling are aligned A unified input channel-Spatial Attention Module (CSAM) enables precise calibration and redundancy suppression of features across scales, adapting to the enhancement requirements of weak features of small infrared targets. Through global pooling and a lightweight MLP, channel weights for features at each scale are adaptively learned for channel attention calibration. Channels enriched by "brightness-variance-gradient" features are prioritized to enhance small target feature channels and suppress background clutter channels. , , , in, For the first Each parallel branch corresponds to a channel-space attention feature at a specific scale.

[0060] To address the limitations of existing attention mechanisms that only enhance features in a single stage and in a single dimension, this invention introduces channel-space joint attention in the two key stages of feature extraction and cross-scale fusion. This allows for two precise calibrations and redundancy suppression of multi-scale features, enhancing the ability to represent shallow small targets and thereby improving the feature differentiation between targets and the background, thus optimizing the overall accuracy of infrared small target detection.

[0061] Step S37: Based on the binary semantic map and channel-spatial attention features at each scale, generate a spatial attention map at each scale, and stitch the spatial attention maps at each scale along the channel dimension to obtain a multi-scale fusion feature map; In this embodiment, the binary semantic map at each scale is... Synchronous upsampling to Size, denoted as A spatial attention map is generated, and higher weights are assigned to suspected target regions (high semantic value, high brightness) to accurately locate the spatial position of small targets. The formula is as follows: , , , in, For the first Spatial attention maps of scales corresponding to each parallel branch For the first Each parallel branch corresponds to a spatial attention weight at a specific scale. This is a channel-dimensional splicing operation.

[0062] Based on the binary semantic map and channel-spatial attention features at each scale, a spatial attention map is generated for each scale. These spatial attention maps are then concatenated along the channel dimension to obtain a multi-scale fused feature map, including: The binary semantic map and channel-spatial attention features at each scale are concatenated along the channel dimension, and then sequentially passed through a 3×3 convolution and a Sigmoid activation function to obtain the spatial attention weights at each scale. The spatial attention map for each scale is obtained by multiplying the channel-spatial attention features of each scale element-wise with the spatial attention weights of that scale. The spatial attention maps at each scale are concatenated along the channel dimension to obtain a multi-scale fused feature map. This multi-scale fused feature map simultaneously contains input preprocessing-enhanced brightness / variance / gradient features, semantic features at each scale, cross-scale interaction features, and precise features after CSAM calibration. It is information-rich with low redundancy and has a high number of channels after concatenation. for: .

[0063] Step S38: Pass the multi-scale fused feature map through the improved three-fusion gate Mamba module to obtain the fused enhanced feature map; The stitched fused feature map Convert to sequence form Infrared segmented position coding is employed, segmenting the data into "suspected target area (high brightness, low variance) / background area." The target area coding weight is increased by 2 times, forming a closed loop with the input preprocessing and the brightness feature logic of the PLSS module. Position code generation: , in, This is the location encoding corresponding to the multi-scale fused feature map. For the suspected target area, This is the background area.

[0064] The sequence characteristics output by the improved three-way fusion gate Mamba module Convert back to 2D feature map To maintain the spatial structural integrity of the features. Sequence features output by the improved three-way fusion gate Mamba module. Channel dimensionality reduction is performed through 1×1 convolution, while simultaneously verifying the consistency of multi-scale semantics—filtering out semantically conflicting clutter features (such as some regions being semantically labeled as targets but with inconsistent brightness features), ensuring the semantic uniformity and accuracy of the fused features. The formula is as follows: , in, To fuse and enhance feature maps, The sequence characteristics output by the improved three-way fusion gate Mamba module.

[0065] The sequence feature map is gated by a three-fold fusion of "input features + infrared segmented position encoding + historical state" to model the long-distance correlation between small targets at different scales and the global background, further filtering discrete background clutter and preventing early small target information from being covered by the background.

[0066] During the training phase, the semantic consistency loss is constrained as follows: , Final output: fused and enhanced feature map The data is fed into the eight-domain clustering module to support the precise positioning of small targets.

[0067] Step S4: Pass the fused and enhanced feature map through an eight-neighborhood clustering module to obtain the infrared small target detection results, including: For fused feature maps Perform sigmoid activation to generate a binary saliency map. A threshold of 0.5 is set—target region value > 0.5, background region value ≤ 0.5—to initially filter semantic target regions. , , in, Location in a binary saliency map The value at that location, This is a binary segmentation mask, where 1 represents pixels in the semantic target region and 0 represents background pixels. Only for binary saliency maps "Semantic target region" The pixels are subjected to eight-neighbor clustering, and brightness feature verification is incorporated to further filter out false targets and improve positioning accuracy. The local brightness mean and variance are calculated for candidate regions to filter out regions whose brightness features do not conform to the characteristics of infrared small targets. For each candidate region, the brightness mean within the semantic target region is calculated. With variance : , , in, For semantic target region, for The total number of pixels, This represents the average local brightness of the semantic target region.

[0068] Set threshold , , retain satisfaction and Candidate regions.

[0069] Perform eight-neighbor clustering on the remaining candidate regions, merge adjacent target pixels to form target candidate boxes. ,in, These are the x and y coordinates of the top left corner of the candidate box, respectively. These are the x and y coordinates of the bottom right corner of the candidate bounding box, respectively. The compactness (aspect ratio < 2) and brightness variance of each cluster region are calculated to filter out stretched clutter regions and false targets with large brightness fluctuations, retaining true small target regions. The aspect ratio of the bounding rectangle is... : , reserve and The region is taken as the real target region.

[0070] Generate the centroid coordinates, bounding rectangle size, and confidence score for each target cluster.

[0071] Centroid coordinates: , Target confidence level: , in, To adjust the parameters, For target candidate boxes Number of pixels included The activation value of the semantic graph. The target region's brightness variance. , These are the x and y coordinates of the centroid of the target cluster, respectively. The target confidence level.

[0072] Output the final infrared small target detection result, including the target's centroid position. Circumscribed rectangle dimensions Confidence level Three core pieces of information meet the engineering application requirements for real-time detection.

[0073] In this embodiment, to facilitate subsequent steps, core experimental parameters are pre-set. The input size of the S6 high-resolution branch is set to 128×128 pixels, and the number of channels is 16; the neighborhood size of the dual-stage channel spatial attention module (CSAM) is set to 3×3, and the weight update threshold is 0.001; the kernel size of each convolutional layer is set to 3×3, the stride is 1, and the padding parameter is "Same". Furthermore, the detection probability Pd (%) is used as the core evaluation metric, defined as: ×100%, in, The number of small infrared targets that are correctly detected. This represents the number of small infrared targets that were not detected correctly.

[0074] It includes three sub-indicators: overall Pd, Pd for micro-targets <4×4 pixels, and Pd for extreme clutter scenes. It also combines false alarm rate Fa (number of false alarms per frame), inter-row union ratio (IoU) (%), and frame rate FPS (frames per second) for comprehensive verification.

[0075] This invention proposes a DAM-Net for infrared small target detection based on brightness perception and three-fusion gate Mamba. Experimental verification is carried out using the publicly available dataset NUAA-SIRST and the self-developed NUDT-SIRST dataset in the field of infrared small target detection. The images in the datasets are uniformly adjusted to 256×256 pixels, including micro-targets <4×4 pixels and complex clutter background scenes.

[0076] (I) Input in this embodiment The input infrared single-channel image is a grayscale image of size 256×256 pixels, with grayscale values ​​normalized to [0,1].

[0077] Parameter settings: 3×3 Gaussian filtering (standard deviation σ=0.8) is used for preprocessing and noise reduction; the input size of the S6 branch is 128×128 pixels and the number of channels is 16; the neighborhood size of the dual-stage channel spatial attention module (CSAM) is set to 3×3 and the weight update threshold is 0.001; the model training batch size is 32, the training epochs are 100, and the patience value of the early stopping strategy is 10.

[0078] (II) Training this deep network model A detection model was constructed using DNA-Net as the base network and embedded with the five innovative modules of this invention. Training was completed on an NVIDIA RTX 4090 GPU using the PyTorch 2.0 framework. From 2880 images expanded from the NUAA-SIRST dataset, the training and test sets were divided in an 8:2 ratio. The training set employed a random flipping and ±15% brightness perturbation strategy to enhance sample diversity.

[0079] The training process employed the AdamW optimizer with an initial learning rate of 0.001, cosine annealing decaying to 1e-6, and a weight decay coefficient of 0.0005. The loss function used was a combination of Focal Loss (α=0.25, γ=2.0) and IoU Loss with a 7:3 weighting to address the imbalanced sample problem. After training, performance was validated on the NUAA-SIRST test set (240 images) and the self-developed NUDT-SIRST independent test set (800 images). Detection probability (Pd), false alarm rate (Fa), intersection-over-union ratio (IoU), and frame rate (FPS) were selected as evaluation metrics.

[0080] Definitions of each indicator: Detection probability (Pd) is the ratio of the number of correctly detected targets to the total number of targets, reflecting the ability to suppress missed detections; False alarm rate (Fa) is the number of falsely detected target points per frame, reflecting the ability to suppress clutter; Intersection over Union (IoU) is the ratio of the area of ​​the intersection and the area of ​​the predicted bounding box to the area of ​​the union of the ground truth bounding box, reflecting the positioning accuracy; Frame rate (FPS) is the number of image frames processed per second, reflecting real-time performance.

[0081] As shown in Table 1, Table 1 presents the performance comparison results of the method of the present invention and the existing mainstream infrared small target detection method HDNet on the NUAA-SIRST test set.

[0082] Table 1

[0083] As can be seen, the method of this invention significantly outperforms existing methods in all core metrics. Pd is improved by 4.3 percentage points compared to the current best model, Fa is reduced by 72.7%, and IoU is improved by 6.9 percentage points, while achieving a frame rate of 45 FPS, balancing detection accuracy and real-time performance, and effectively reducing missed detections and false alarms. Furthermore, in the self-developed NUDT-SIRST extreme scenario test set, the method of this invention still achieves a Pd of 96.2% and a Fa of 0.5 per frame, with a performance degradation of only 2.3%, far lower than the comparative model. This demonstrates that the method of this invention possesses strong model robustness.

[0084] This second embodiment provides an infrared small target detection system, including: The feature extraction module is used to take the preprocessed infrared single-channel grayscale image as the input image of the feature extraction unit, and use the feature extraction unit to extract features at different scales. The feature fusion module is used to pass features of different scales through the feature fusion unit to obtain a fused and enhanced feature map, including: Each scale feature is extracted by using depthwise separable convolution and 1×1 convolution; Based on the local brightness mean and local brightness variance of each scale feature, a brightness mask for each scale is generated. After multiplying the spatial semantic features of each scale element-wise with the brightness mask, the binary semantic map of each scale is obtained by passing it through the Sigmoid activation function. Based on the binary semantic graph at each scale, obtain the global semantic confidence weight for each scale; Each scale feature is multiplied element-wise by the binary semantic graph of that scale, and then multiplied element-wise by the global semantic confidence weight of that scale to obtain the weighted feature of each scale. After upsampling the weighted features of each scale to the maximum resolution scale among the features of each scale, the channel-spatial attention module is used to obtain the channel-spatial attention features of each scale. Based on the binary semantic map and channel-space attention features at each scale, a spatial attention map at each scale is generated. The spatial attention maps at each scale are then spliced ​​along the channel dimension to obtain a multi-scale fusion feature map. The multi-scale fused feature map is passed through the improved three-fusion gate Mamba module to obtain the fused enhanced feature map; The detection module is used to obtain infrared small target detection results by passing the fused and enhanced feature map through the eight-neighbor clustering module.

[0085] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0086] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0087] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0088] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0089] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for detecting small infrared targets, characterized in that, include: The preprocessed infrared single-channel grayscale image is used as the input image of the feature extraction unit, and features at different scales are extracted using the feature extraction unit. Features at different scales are processed by a feature fusion unit to obtain a fused and enhanced feature map, including: Each scale feature is extracted by using depthwise separable convolution and 1×1 convolution; Based on the local brightness mean and local brightness variance of each scale feature, a brightness mask for each scale is generated. After multiplying the spatial semantic features of each scale element-wise with the brightness mask, the binary semantic map of each scale is obtained by passing it through the Sigmoid activation function. Based on the binary semantic graph at each scale, obtain the global semantic confidence weight for each scale; Each scale feature is multiplied element-wise by the binary semantic graph of that scale, and then multiplied element-wise by the global semantic confidence weight of that scale to obtain the weighted feature of each scale. After upsampling the weighted features of each scale to the maximum resolution scale among the features of each scale, the channel-spatial attention module is used to obtain the channel-spatial attention features of each scale. Based on the binary semantic map and channel-space attention features at each scale, a spatial attention map at each scale is generated. The spatial attention maps at each scale are then spliced ​​along the channel dimension to obtain a multi-scale fusion feature map. The multi-scale fused feature map is passed through the improved three-fusion gate Mamba module to obtain the fused enhanced feature map; The fused and enhanced feature maps are then processed through an eight-domain clustering module to obtain infrared small target detection results.

2. The infrared small target detection method according to claim 1, characterized in that, The process of acquiring the preprocessed infrared single-channel grayscale image includes: The infrared single-channel grayscale image is normalized to obtain the normalized infrared single-channel grayscale image. A 3×3 Gaussian filter was used to denoise the normalized infrared single-channel grayscale image to obtain a preprocessed infrared single-channel grayscale image.

3. The infrared small target detection method according to claim 1, characterized in that, Methods for optimizing preprocessed infrared single-channel grayscale images to obtain enhanced input images for feature extraction units include: The preprocessed infrared single-channel grayscale image is used as the original brightness feature; A 3×3 sliding window is used to traverse all pixels of the preprocessed infrared single-channel grayscale image, and the local brightness variance is calculated to obtain the local variance features. Sobel gradient calculation is performed on the preprocessed infrared single-channel grayscale image to calculate its gradient in the horizontal and vertical directions respectively. The brightness gradient features are obtained by fusing the horizontal and vertical gradients of the preprocessed infrared single-channel grayscale image. The original brightness features, local variance features, and brightness gradient features are concatenated along the channel dimension to obtain the enhanced input image of the infrared small target detection model.

4. The infrared small target detection method according to claim 1, characterized in that, The feature extraction unit consists of multiple parallel branches, each of which is a densely nested attention network that extracts features at different scales. The shallow branch with the highest resolution is half the size of the input image and has 16 channels.

5. The infrared small target detection method according to claim 4, characterized in that, The densely nested attention network consists of multiple layers connected in sequence. Each layer includes a convolutional layer and a channel-space attention module connected in sequence. Dense skip connections and bidirectional interactive skip connections are used between the layers.

6. The infrared small target detection method according to claim 5, characterized in that, The last convolutional layer in the deepest branch with the lowest resolution is replaced with the improved three-gate Mamba module to obtain the target feature extraction unit.

7. The infrared small target detection method according to claim 6, characterized in that, Methods for obtaining the output feature map of the improved three-gate Mamba module by passing the input feature map through the improved three-gate Mamba module include: The input feature map is flattened in the spatial dimension to obtain a sequence of input feature maps; infrared segmentation position encoding is applied to the sequence of multi-scale fusion features to obtain position codes; For each sequence position in the input feature map in sequence form, perform a linear transformation on the input feature value, position code, and historical state information of the previous sequence position to obtain the linearly transformed input feature value, position code, and historical state information of the previous sequence position. The updated input feature values ​​at the current sequence position are passed sequentially through a multilayer perceptron and a softmax activation function to obtain the adaptive fusion weights at the current sequence position. Based on the adaptive fusion weights of the current sequence position, the input feature value position encoding after linear transformation of the current sequence position and the historical state information of the previous sequence position are weighted and summed, and then the forgetting gate vector of the current sequence position is obtained by passing the Sigmoid activation function. Based on the state decay coefficient, the forgetting gate vector of the current sequence position, the input feature value after linear transformation of the current sequence position, and the historical state information after linear transformation of the previous sequence position, the historical state information of the current sequence position is obtained. Based on the feature value after linear transformation of the current sequence position and historical state information, obtain the final feature value of the current sequence position; After all positions in the entire sequence have been processed, the final feature values ​​output from each sequence position are concatenated in the original sequence order and reconstructed into a two-dimensional feature form, resulting in the output feature map of the improved three-way Mamba module.

8. The infrared small target detection method according to claim 7, characterized in that, The formula for obtaining the historical state information of the current sequence position based on the state decay coefficient, the forgetting gate vector of the current sequence position, the input feature value after linear transformation of the current sequence position, and the historical state information after linear transformation of the previous sequence position is as follows: ， in, Current sequence position Historical status information Position of the previous sequence Historical state information after linear transformation The state decay coefficient, Current sequence position The forgetting gate vector, The input feature value is the linear transformation of the current sequence position.

9. The infrared small target detection method according to claim 1, characterized in that, The method for obtaining the global semantic confidence weight for each scale based on the binary semantic graph at each scale includes: For each scale of the binary semantic map, pixels with values ​​greater than a set threshold are taken as target pixels, and the ratio of the number of target pixels to the total number of pixels in the binary semantic map is used as the global semantic confidence weight for the corresponding scale.

10. An infrared small target detection system, characterized in that, include: The feature extraction module is used to take the preprocessed infrared single-channel grayscale image as the input image of the feature extraction unit, and use the feature extraction unit to extract features at different scales. The feature fusion module is used to pass features of different scales through the feature fusion unit to obtain a fused and enhanced feature map, including: Each scale feature is extracted by using depthwise separable convolution and 1×1 convolution; Based on the local brightness mean and local brightness variance of each scale feature, a brightness mask for each scale is generated. After multiplying the spatial semantic features of each scale element-wise with the brightness mask, the binary semantic map of each scale is obtained by passing it through the Sigmoid activation function. Based on the binary semantic graph at each scale, obtain the global semantic confidence weight for each scale; Each scale feature is multiplied element-wise by the binary semantic graph of that scale, and then multiplied element-wise by the global semantic confidence weight of that scale to obtain the weighted feature of each scale. After upsampling the weighted features of each scale to the maximum resolution scale among the features of each scale, the channel-spatial attention module is used to obtain the channel-spatial attention features of each scale. Based on the binary semantic map and channel-space attention features at each scale, a spatial attention map at each scale is generated. The spatial attention maps at each scale are then spliced ​​along the channel dimension to obtain a multi-scale fusion feature map. The multi-scale fused feature map is passed through the improved three-fusion gate Mamba module to obtain the fused enhanced feature map; The detection module is used to obtain infrared small target detection results by passing the fused and enhanced feature map through the eight-neighbor clustering module.

Images