An improved infrared small target detection method and system based on visual state space model

Abstract

The present invention belongs to the field of visual detection technology, and specifically relates to a method and system for detecting small infrared targets based on an improved visual state space model. The method comprises: preprocessing the collected original infrared image and then inputting it into a pre-established and trained improved model to achieve real-time remote sensing recognition of small infrared targets. The improved model includes: a feature extraction backbone network, improved based on Mamba, for extracting small target features from the input image; a feature fusion head network for multi-scale fusion of the different-scale features extracted by the backbone network to generate different-scale feature layers; and a target detection network for performing target classification and regression on the different-scale feature layers and outputting detection results. Based on the visual Mamba, the present invention improves the backbone network structure to effectively improve the accuracy of small target recognition, targeting the characteristics of small infrared targets.

Description

Technical Field

[0001] The present invention belongs to the technical field of visual detection, and in particular relates to an improved infrared small target detection method and system based on a visual state space model. Background Art

[0002] In recent years, small target detection technology has demonstrated its broad application potential in a variety of key areas, including but not limited to observing wildlife migration, detecting crop pests and diseases, detecting urban vehicles, and detecting small military targets. However, the inherent complexity of remote sensing images, such as the small and dense size of targets, complex background environments, and numerous interference factors, significantly hinders the accuracy of target detection and recognition.

[0003] Furthermore, with the advancement of remote sensing imaging equipment, computational efficiency for high-resolution detection has become a focus of attention. As the input image resolution increases, the computational complexity of the convolutional layers in models based on convolutional architectures increases significantly, requiring deeper network structures to maintain performance. The computational complexity of models based on the Transformer architecture increases quadratically with the number of pixels in the input image. When the image resolution is too high, the model cannot function properly due to insufficient memory, limiting its ability to process high-resolution images.

[0004] Introducing a linearly complex state-space model (Mamba) into vision tasks can address computational constraints in high-resolution images. However, directly using Mamba for small object detection still suffers from low detection rates. This is primarily due to the small size of the objects and the complex backgrounds. Accurately identifying these small objects requires not only Mamba's global attention mechanism but also the ability to capture fine local features and resist background interference.

[0005] Therefore, how to improve the accuracy of small target recognition in high-resolution remote sensing images has become an important issue that needs to be solved urgently in the current research field. Summary of the Invention

[0006] The purpose of the present invention is to overcome the defects of the prior art and propose an improved infrared small target detection method and system based on a visual state space model.

[0007] In view of this, the present invention proposes an infrared small target detection method based on an improved visual state space model, comprising:

[0008] After pre-processing the collected original infrared images, the images are input into the pre-established and trained improved model to achieve real-time remote sensing recognition of small infrared targets.

[0009] The improved model includes: a feature extraction backbone network, a feature fusion head network and a target detection network, wherein,

[0010] The feature extraction backbone network is improved based on Mamba and is used to extract small object features in the input image;

[0011] The feature fusion head network is used to perform multi-scale fusion on the different-scale features extracted by the backbone network to generate feature layers of different scales;

[0012] The target detection network is used to perform target classification and regression on feature layers of different scales and output detection results.

[0013] Preferably, the feature extraction backbone network includes: 1 Stem Layer and 4 Stage Layers, the 4 Stage Layers are: Stage Layer 1, Stage Layer 2, Stage Layer 3 and Stage Layer 4; wherein,

[0014] The Stem Layer is a convolution layer with a convolution kernel size of 3 and a stride of 2, which is used to preliminarily extract low-level features of edges and textures in the input image;

[0015] The Stage Layer 1 consists of a convolution layer with a kernel size of 3 and a stride of 2 and an ESTVSS module, which is used to mine more complex features in the image based on the preliminary feature extraction;

[0016] The Stage Layer 2 includes a VisionClueMerge module and an ESTVSS module connected in sequence. The VisionClueMerge module is used to perform feature recombination and a series of linear transformation, normalization and activation operations on the input features. The output of the ESTVSS module is a P3 large-scale feature map.

[0017] The Stage Layer 3 includes a VisionClueMerge module and an ESTVSS module connected in sequence, and the output of the ESTVSS module is a P4 medium-scale feature map;

[0018] The Stage Layer 4 includes a VisionClueMerge module, an ESTVSS module, and an SPPF module connected in sequence; the SPPF module is used to perform fast spatial pyramid pooling and feature splicing on the input feature map, using a small pooling kernel of size 5 to perform two pooling operations instead of a single large kernel pooling operation to reduce the amount of computation, and outputting a P5 small-scale feature map.

[0019] Preferably, the ESTVSS modules in the four Stage Layers include: SS2D module, ESTD module and CARG module; wherein,

[0020] The SS2D module is used to effectively extend the state-space model's ability to process one-dimensional sequential data to two-dimensional spatial visual data, and achieves feature extraction with linear time complexity by traversing the image along four scanning paths;

[0021] The ESTD module is used to enhance local attention and thus optimize the size of the receptive field, so that the model can better capture the fine local details of small objects;

[0022] The CARG module is used to enhance channel and spatial attention capabilities, enabling the model to better distinguish complex backgrounds and targets.

[0023] Preferably, the output X of the ESTD module ESTD for:

[0024] X ESTD =Conv 1×1 (σ GELU (Conv 1×1 (SE(BN(Conv 1×1 (X proj ))))))

[0025] X proj =σ SiLU (BN(Conv 1×1 (X)))

[0026] Among them, X is the input tensor of the ESTVSS module; Conv 1×1 is a convolution layer with a convolution kernel of 1; BN is batch normalization; σ SiLU is the SiLU activation function; SE is the compression and excitation module; σ GELU is the GELU activation function.

[0027] Preferably, the CARG module includes a Channel Attention submodule and a Spatial Attention submodule, wherein:

[0028] The Channel Attention submodule is used to enhance channels containing key features while suppressing channels irrelevant to the features of the small target itself, reducing the interference of irrelevant information and enabling the model to focus on the effective features of the small target;

[0029] The Spatial Attention submodule is used to enhance the spatial location information features of small targets, reduce the weight of the surrounding background area, shift the model's attention from the background area to the small target area, and weaken the interference of the background on small target detection.

[0030] Preferably, the output x of the Channel Attention submodule is channelattention for:

[0031] x channelattention =σ sig (x out_avgpool +x out_maxpool )

[0032] x out_maxpool =Conv 1×1 (σ ReLU (Conv 1×1 (f AMP (x DW_conv ))))

[0033] x out_avgpool =Conv 1×1 (σReLU(Conv 1×1 (f AAP (x DW_conv ))))

[0034] Among them, x DW_conv is a depth-wise separable convolution; f AMP is the adaptive maximum pooling layer; f AAP is the adaptive average pooling layer; Conv 1×1 is a convolution layer with a convolution kernel of 1; σ ReLU is the ReLU activation function; σ Sig is the Sigmoid activation function.

[0035] Preferably, the output x of the Spatial Attention submodule is SA_out for:

[0036] x SA_out =σ Sig (Conv 1×1 (Cat[x mean ,x max ]))

[0037] x mean =ChannelMean(x CA_out )

[0038] x max =ChannelMax(x CA_out )

[0039] x CA_out =x channelattention ×x DW_conv

[0040] Among them, ChannelMean is the channel mean; ChannelLmax is the channel maximum value; Cat is the channel splicing and merging; Conv 1×1 It is a convolution layer with a convolution kernel of 1.

[0041] On the other hand, the present invention provides an infrared small target detection system based on an improved visual state space model, comprising:

[0042] A preprocessing module is used to preprocess the collected original infrared images;

[0043] The detection output module is used to input the pre-processed image into the pre-established and trained improved model to achieve remote sensing real-time recognition of small infrared targets;

[0044] The improved model includes: a feature extraction backbone network, a feature fusion head network and a target detection network, wherein,

[0045] The feature extraction backbone network is improved based on Mamba and is used to extract small object features in the input image;

[0046] The feature fusion head network is used to perform multi-scale fusion on the different-scale features extracted by the backbone network to generate feature layers of different scales;

[0047] The target detection network is used to perform target classification and regression on feature layers of different scales and output detection results.

[0048] Compared with the prior art, the advantages of the present invention are:

[0049] 1. The improved infrared small target detection method based on the visual state space model of the present invention has the advantages of not being limited by high-resolution pixels, being lightweight, and having high small target recognition accuracy.

[0050] 2. Compared with the original visual state space model, the improved infrared small target detection method based on the visual state space model of the present invention optimizes the local detail capture capability of the model by designing the ESTD module, making it more suitable for small target detection; and enhances the attention to the spatial position and feature channel information of small targets by designing the CARG module to deal with the problem of complex background interference.

[0051] 3. Compared with the CNN architecture model, the improved infrared small target detection method based on the visual state space model of the present invention has the ability of global context reasoning; compared with the Transformer architecture model, our method has linear complexity. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] FIG1 is a block diagram comparison of a feature extraction backbone network of the present invention and a network before improvement, wherein FIG1(a) is a feature extraction backbone network before improvement, and FIG1(b) is a feature extraction backbone network of the present invention;

[0053] Figure 2 This is a structural diagram of the ESTVSS module of the present invention;

[0054] Figure 3 This is a structural diagram of the SS2D module in the ESTVSS module of the present invention;

[0055] Figure 4 This is a structural diagram of the VisionClueMerge module in the feature extraction network of the present invention;

[0056] FIG5 is a structural diagram of the ESTD module and the CARG module of the present invention, wherein FIG5(a) is the ESTD module and FIG5(b) is the CARG module;

[0057] Figure 6 This is a rendering of the ESTD module of the present invention;

[0058] Figure 7 This is a rendering of the CARG module of the present invention;

[0059] Figure 8 It is a structural diagram of the improved model of the present invention. DETAILED DESCRIPTION

[0060] The present invention proposes an infrared small target detection method based on an improved visual state space model, comprising:

[0061] After pre-processing the collected original infrared images, the images are input into the pre-established and trained improved model to achieve real-time remote sensing recognition of small infrared targets.

[0062] Improved models include:

[0063] Feature extraction backbone network, improved based on Mamba, used to extract small object features in the input image;

[0064] The feature fusion head network is used to perform multi-scale fusion on the different-scale features extracted by the backbone network to generate feature layers of different scales;

[0065] The target detection network is used to perform target classification and regression on feature layers of different scales and output detection results.

[0066] The technical solution of the present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0067] Example 1

[0068] Embodiment 1 of the present invention proposes an infrared small target detection method based on an improved visual state space model. The method is a Mamba neural network model improved according to the characteristics of small targets.

[0069] The improved model includes: a visual state space feature extraction backbone network based on Mamba improvement, a feature fusion head network based on YOLOv8, and a target detection network;

[0070] FIG1 is a comparison of the feature extraction backbone network of the present invention and the network block diagram before improvement. FIG1( a ) is the feature extraction backbone network before improvement, and FIG1( b ) is the feature extraction backbone network of the present invention.

[0071] An improved feature extraction backbone network based on Mamba is used to extract features of small targets in the input image. The original infrared images collected by the drone undergo preprocessing operations such as cropping and flipping before being fed into the improved visual state space backbone network. This backbone network enhances its ability to extract small target features by adding the designed ESTD and CARG modules to the ESTVSS module.

[0072] The improved feature extraction backbone network consists of 1 Stem Layer and 4 Stage Layers, which are respectively recorded as Stage Layer 1, Stage Layer 2, Stage Layer 3 and Stage Layer 4; among them,

[0073] The Stem layer is a convolutional layer with a kernel size of 3 and a stride of 2. It is used to initially extract low-level features such as edges and textures from the input image. The stride of 2 is set to reduce the amount of computation while retaining important information. The height and width of the feature map output after the Stem layer are half of the original image.

[0074] Stage Layer 1 consists of a convolutional layer with a kernel size of 3 and a stride of 2, and an ESTVSS module. This layer is used to mine more complex features within the image based on the initial feature extraction. The length and width of the output feature map after the convolutional layer are one-quarter the original image. This feature map is then fed into the ESTVSS module to extract shape features for small objects.

[0075] Stage Layer 2 consists of the VisionClueMerge module and the ESTVSS module. The VisionClueMerge module is used to reorganize input features and perform a series of linear transformations, normalization, and activation operations. The output of this module serves as the input to the subsequent ESTVSS module and continues to participate in the feature extraction process of the entire network. The output of the ESTVSS module in Stage Layer 2 is defined as the P3 large-scale feature map, which is fed into the YOLOv8 head network for feature fusion.

[0076] Stage Layer 3 consists of the VisionClueMerge module and the ESTVSS module. The output of the ESTVSS module in Stage Layer 3 is defined as the medium-scale feature map in P4, which is fed into the YOLOv8 head network for feature fusion.

[0077] Stage Layer 4 consists of the VisionClueMerge module, the ESTVSS module, and the SPPF module. The SPPF module performs fast spatial pyramid pooling and feature concatenation on the input feature map. It uses a small pooling kernel of size 5 for two pooling passes instead of a single large kernel pooling pass to reduce computational complexity. The output of the SPPF module in Stage Layer 4 is defined as a P5 small-scale feature map, which is fed into the YOLOv8 head network for feature fusion.

[0078] The improved ESTVSS module of the feature extraction backbone network includes SS2D module, ESTD module, CARG module and other modules; among them,

[0079] The SS2D module effectively extends the state-space model's ability to process one-dimensional sequential data to two-dimensional visual data. By traversing the image along four scanning paths, a feature extraction method with linear time complexity is constructed.

[0080] The ESTD module is used to enhance local attention and thus optimize the size of the receptive field, enabling the model to better capture the fine local details of small targets.

[0081] The CARG module is used to enhance channel and spatial attention capabilities, enabling the model to better distinguish complex backgrounds and targets.

[0082] The specific implementation of the ESTD module of the ESTVSS module can be expressed as

[0083] X proj =σ SiLU (BN(Conv 1×1 (X)))

[0084] X ESTD=Conv 1×1 (σ GELU (Conv 1×1 (SE(BN(Conv 1×1 (X proj ))))))

[0085] Among them, X is the input tensor of the ESTVSS module; Conv 1×1 is a convolution layer with a convolution kernel of 1; BN is batch normalization; σ SiLU is the SiLU activation function; SE is the compression and excitation module; σ GELU is the GELU activation function.

[0086] The CARG module of the ESTVSS module consists of the Channel Attention submodule and the Spatial Attention submodule;

[0087] The Channel Attention submodule is used to enhance channels containing key features while suppressing channels that are irrelevant to the features of the small target itself, thereby reducing the interference of irrelevant information such as background, allowing the model to focus on the effective features of the small target.

[0088] The Spatial Attention submodule is used to enhance the spatial location information features of small targets and reduce the weight of the surrounding background areas, thereby shifting the model's attention from the background areas to the small target areas and weakening the interference of the background on small target detection.

[0089] The specific implementation of the Channel Attention submodule in the CARG module can be expressed as:

[0090] x out_maxpool =Conv 1×1 (σ ReLU (Conv 1×1 (f AMP (x DW_conv ))))

[0091] x out_avgpool =Conv 1×1 (σ ReLU (Conv 1×1 (f AAP (x DW_conv ))))

[0092] x channelattention =σ Sig (x out_avgpool +x out_maxpool )

[0093] Among them, x DW_convis a depth-wise separable convolution; f AMP is the adaptive maximum pooling layer; f AAP is the adaptive average pooling layer; Conv 1×1 is a convolution layer with a convolution kernel of 1; σ ReLU is the ReLU activation function; σ Sig is the Sigmoid activation function.

[0094] The specific implementation of the Spatial Attention submodule in the CARG module can be expressed as:

[0095] x mean =ChannelMean(x CA_out )

[0096] x max =ChannelMax(x CA_out )

[0097] x SA_out =σ Sig (Conv 1×1 (Cat[x mean ,x max ]))

[0098] x CA_out =x channelattention ×x DW_conv

[0099] Among them, ChannelMean is the channel mean; ChannelMax is the channel maximum value; Cat is the channel splicing and merging; Conv 1×1 It is a convolution layer with a convolution kernel of 1.

[0100] The head network based on YOLOv8 is used to perform multi-scale fusion of the different scale features extracted by the backbone network to generate P3, P4 and P5 different scale feature layers;

[0101] The target detection network is used to send the P3, P4, and P5 feature layers to the detection head network for target classification and regression and output the detection results.

[0102] The specific steps are as follows: first, the original infrared image captured by the drone is preprocessed; then, the processed image is input into the improved visual state space backbone network for small target feature extraction. The improved backbone network uses Mamba's global attention mechanism to capture contextual information dependencies within the image, enhances the model's ability to extract local detail features through the ESTD module, and enhances the spatial position and feature channel information of small targets through the CARG module to address the problem of complex background interference. The feature map obtained from the feature extraction is then fed into the head network for multi-scale fusion and generates feature layers of different scales. Finally, the feature layers of different scales are fed into the target detection network for target classification and regression, and the detection results are output. This realizes the detection function of remote sensing real-time recognition of small infrared targets.

[0103] Figure 1(b) shows the structure of the feature extraction backbone network of the present invention, which includes a convolutional layer, an ESTVSS module, a VisionClueMerge module, and an SPPF module; wherein,

[0104] The convolutional layer is used to initially extract primary texture information;

[0105] The ESTVSS module is used to capture the local detail texture and global image features of the target; it consists of the ESTD module, normalization layer, SS2D module and CARG module;

[0106] VisionClueMerge module (such as Figure 4 As shown in Figure 2, it is used to perform feature reorganization and a series of linear transformation, normalization and activation operations on the input features. The output of this module will serve as the input of the subsequent ESTVSS module and continue to participate in the feature extraction process of the entire network.

[0107] The SPPF module is used to perform fast spatial pyramid pooling operations and feature splicing on the input feature map.

[0108] The use of the ESTVSS module can capture effective information more comprehensively, thereby improving the accuracy of small target detection.

[0109] Figure 2 This is the ESTVSS module structure diagram of the present invention, including SS2D module, ESTD module and CARG module and other modules; wherein,

[0110] The SS2D module (such as Figure 3 The ability of the state-space model to process one-dimensional sequential data is effectively extended to two-dimensional spatial visual data. By traversing the image along four scanning paths, a feature extraction method with linear time complexity is constructed.

[0111] FIG5 is a structural diagram of the ESTD module and the CARG module of the present invention.

[0112] The ESTD module (as shown in Figure 5(a)) is used to enhance local attention and thus optimize the size of the receptive field, so that the model can better capture the fine local details of small targets.

[0113] The CARG module (as shown in FIG5( b )) is used to enhance channel and spatial attention capabilities, so that the model can better distinguish complex backgrounds and targets.

[0114] Figure 6 This is the rendering of the ESTD module of the present invention. By comparing the visualization effects before and after adding the ESTD module, it is shown that the ESTD module designed by the present invention can accurately capture the local details of small targets.

[0115] Figure 7 This is a rendering of the CARG module of the present invention. Comparing the visualization effects before and after adding the CARG module shows that the CARG module designed by the present invention can better distinguish complex backgrounds and targets, thereby improving the detection accuracy of small targets.

[0116] Figure 8 This is a structural diagram of the improved model of the present invention.

[0117] Table 1 shows the experimental results of the present invention and other models. By comparing the accuracy index mAP, it is shown that the infrared small target detection method designed by the present invention based on the improved visual state space model can effectively improve the detection and recognition accuracy of small targets.

[0118] Table 1 Comparative experimental results of the present invention and other models

[0119]

[0120] Example 2

[0121] Embodiment 2 of the present invention provides an infrared small target detection system based on an improved visual state space model, which is implemented based on the method of embodiment 1 and includes:

[0122] An image preprocessing module is used to preprocess the collected original infrared image; in one embodiment, the entire image is divided into blocks;

[0123] The detection output module is used to input the pre-processed image into the pre-established and trained improved model to achieve remote sensing real-time recognition of small infrared targets;

[0124] The improved model includes: a feature extraction backbone network, a feature fusion head network and a target detection network, wherein,

[0125] The feature extraction backbone network is improved based on Mamba and is used to extract small object features in the input image;

[0126] The feature fusion head network is used to perform multi-scale fusion on the different-scale features extracted by the backbone network to generate feature layers of different scales;

[0127] The target detection network is used to perform target classification and regression on feature layers of different scales and output detection results.

[0128] Finally, it should be noted that the above embodiments are intended only to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the embodiments, it should be understood by those skilled in the art that modifications or equivalent substitutions to the technical solutions of the present invention do not depart from the spirit and scope of the technical solutions of the present invention and are intended to be encompassed by the claims of the present invention.

Claims

1. An improved infrared small target detection method based on a visual state space model, comprising: After pre-processing the collected original infrared images, the images are input into the pre-established and trained improved model to achieve real-time remote sensing recognition of small infrared targets. The improved model includes: a feature extraction backbone network, a feature fusion head network and a target detection network, wherein, The feature extraction backbone network is improved based on Mamba and is used to extract small object features in the input image; The feature fusion head network is used to perform multi-scale fusion on the different-scale features extracted by the backbone network to generate feature layers of different scales; The target detection network is used to perform target classification and regression on feature layers of different scales and output detection results; The feature extraction backbone network includes: 1 Stem Layer and 4 Stage Layers, the 4 Stage Layers are: Stage Layer 1, Stage Layer 2, Stage Layer 3 and Stage Layer 4; wherein, The Stem Layer is a convolution layer with a convolution kernel size of 3 and a stride of 2, which is used to preliminarily extract low-level features of edges and textures in the input image; The Stage Layer 1 consists of a convolution layer with a kernel size of 3 and a stride of 2 and an ESTVSS module, which is used to mine more complex features in the image based on the preliminary feature extraction; The Stage Layer 2 includes a VisionClueMerge module and an ESTVSS module connected in sequence. The VisionClueMerge module is used to perform feature recombination and a series of linear transformation, normalization and activation operations on the input features. The output of the ESTVSS module is a P3 large-scale feature map. The Stage Layer 3 includes a VisionClueMerge module and an ESTVSS module connected in sequence, and the output of the ESTVSS module is a P4 medium-scale feature map; The Stage Layer 4 includes a VisionClueMerge module, an ESTVSS module, and an SPPF module connected in sequence; the SPPF module is used to perform fast spatial pyramid pooling and feature splicing on the input feature map, using a small pooling kernel of size 5 to perform two pooling operations instead of a single large kernel pooling operation to reduce the amount of calculation, and outputting a P5 small-scale feature map; The ESTVSS modules in the four Stage Layers include: SS2D module, ESTD module and CARG module; wherein, The SS2D module is used to effectively extend the state-space model's ability to process one-dimensional sequential data to two-dimensional spatial visual data, and achieves feature extraction with linear time complexity by traversing the image along four scanning paths; The ESTD module is used to enhance local attention and thus optimize the size of the receptive field, so that the model can better capture the fine local details of small objects; The CARG module is used to enhance channel and spatial attention capabilities, enabling the model to better distinguish complex backgrounds and targets.

2. The infrared small target detection method based on the improved visual state space model according to claim 1 is characterized in that: The output X of the ESTD module ESTD for: X ESTD =Conv 1×1 (σ GELU (Conv 1×1 (SE(BN(Conv 1×1 (X proj )))))) X proj =σ SiLU (BN(Conv 1×1 (X))) Among them, X is the input tensor of the ESTVSS module, Conv 1×1 is a convolution layer with a convolution kernel of 1, BN is batch normalization, σ SiLU is the SiLU activation function, SE is the compression and excitation module, σ GELU is the GELU activation function.

3. The infrared small target detection method based on the improved visual state space model according to claim 1 is characterized in that: The CARG module includes a Channel Attention submodule and a Spatial Attention submodule, wherein: The Channel Attention submodule is used to enhance channels containing key features while suppressing channels irrelevant to the features of the small target itself, reducing the interference of irrelevant information and enabling the model to focus on the effective features of the small target; The Spatial Attention submodule is used to enhance the spatial location information features of small targets, reduce the weight of the surrounding background area, shift the model's attention from the background area to the small target area, and weaken the interference of the background on small target detection.

4. The infrared small target detection method based on the improved visual state space model according to claim 3 is characterized in that: The output x of the Channel Attention submodule channelattention for: x channelattention =s Sig (x out_avgpool +x out_maxpool ) x out_maxpool =Conv 1×1 (σ ReLU (Conv 1×1 (f AMP (x DW_conv )))) x out_avgpool =Conv 1×1 (σ ReLU (Conv 1×1 (f AAP (x DW_conv )))) Among them, x DW_conv is a depth-wise separable convolution, f AMP is the adaptive maximum pooling layer, f AAP is the adaptive average pooling layer, Conv 1×1 is a convolution layer with a convolution kernel of 1, σ ReLU is the ReLU activation function, σ Sig is the Sigmoid activation function.

5. The infrared small target detection method based on the improved visual state space model according to claim 4 is characterized in that: The output x of the Spatial Attention submodule SA_out for: x SA_out =σ Sig (Conv 1×1 (Cat[x mean ,x max ])) x mean =ChannelMean(x CA_out ) x max =ChannelMax(x CA_out ) x CA_out =x channelattention ×x DW_conv Among them, ChannelMean is the channel mean, ChannelMax is the channel maximum value, Cat is the channel splicing and merging, Conv 1×1 It is a convolution layer with a convolution kernel of 1.

6. A system based on the improved infrared small target detection method based on the visual state space model according to claim 1, characterized in that: include: A preprocessing module is used to preprocess the collected original infrared images; and The detection output module is used to input the pre-processed image into the pre-established and trained improved model to achieve remote sensing real-time recognition of small infrared targets; The improved model includes: a feature extraction backbone network, a feature fusion head network and a target detection network, wherein, The feature extraction backbone network is improved based on Mamba and is used to extract small object features in the input image; The feature fusion head network is used to perform multi-scale fusion on the different-scale features extracted by the backbone network to generate feature layers of different scales; The target detection network is used to perform target classification and regression on feature layers of different scales and output detection results.

Images