A dual-branch infrared small target recognition method based on selective state space

Abstract

This invention discloses a dual-branch infrared small target recognition method based on selective state space. It collects and divides an infrared small target dataset, uses a ResNet-50 backbone network to extract features from the data, and then inputs these features into a Region of Interest (RPN) module and a Key Attention Aggregation (KAAM) module to further extract semantic representations and location information, which are then fused to achieve accurate recognition of small targets in complex backgrounds. Furthermore, a target discrimination network is used, and an SS2D spatial selective modeling module is introduced to deeply extract global contextual information and dynamically model the spatiotemporal dependency between the target and the background. This invention effectively overcomes the limitations of traditional methods in recognizing complex background interference and target size variations. Simultaneously, an optimization strategy based on contrast constraints is designed during the network model training process to enhance the network's accuracy in identifying and segmenting small targets, achieving highly robust small target recognition performance.

Description

Technical Field

[0001] This invention belongs to the fields of deep learning, infrared imaging, and small target detection technology, and relates to a dual-branch infrared small target recognition method based on selective state space. Background Technology

[0002] Infrared small target recognition technology is a core challenge for computer vision in reconnaissance, precision guidance, and other fields. Its core difficulty stems from the "weak, small, and sparse" characteristics of the targets: "Weak": Targets have low thermal radiation energy and are easily obscured by complex backgrounds (such as clouds and waves) and sensor noise; "Small": Targets typically occupy only 1-10 pixels, lacking shape and texture features; "Sparse": Target distribution is highly random, leading to extreme imbalance between positive and negative samples. Traditional methods are mainly model-driven, including filtering-based methods (such as Gaussian-Laplacian operators), local contrast methods (such as LCM and MPCM), and low-rank sparse decomposition methods (such as IPI and RIPT), which suppress background clutter through manually designed features. However, these methods have high false alarm rates in dynamic backgrounds such as moving clouds and rely on static assumptions, making them difficult to handle complex scenarios with missing temporal information. In recent years, deep learning has made significant progress in infrared small target recognition, giving rise to a series of innovative network architectures to address the core challenges of "weak, small, and sparse" targets. For example, HCF-Net enhances global context awareness while preserving target details through a parallelized block-aware attention module and a multi-expansion channel refinement mechanism, significantly improving detection accuracy in complex backgrounds. ISTDU-Net, based on an improved U-Net architecture, introduces feature map grouping awareness and fully connected skip layers to effectively suppress noise interference and enhance the feature representation of small targets. Addressing real-time requirements, YOLO-MST combines super-resolution reconstruction with a multi-scale dynamic detection head, optimizing detection speed through dilated convolutional feature aggregation, achieving efficient inference in embedded scenarios. Furthermore, ISNet solves the challenge of target shape reconstruction under low signal-to-noise ratio by using Taylor finite difference edge blocks and a bidirectional attention aggregation mechanism, while SuperYOLO innovatively integrates infrared and visible light modal data, utilizing multi-modal complementary information to enhance the feature representation of small targets. These networks, through multi-scale feature fusion, dynamic attention mechanisms, and lightweight design, are gradually overcoming the bottlenecks of traditional methods such as dynamic background interference, deep feature degradation, and insufficient real-time performance, driving the continuous evolution of infrared small target detection technology towards higher accuracy, higher robustness, and lower power consumption.

[0003] Current infrared small target detection methods primarily focus on feature extraction from the target's core region. However, existing technologies still face two major challenges: firstly, insufficient detection accuracy for small targets in complex background interference; and secondly, weak distinguishability between targets and noise leading to a high false alarm rate. These problems blur the target's edge contours, resulting in frequent missed and false detections. Therefore, effectively fusing multi-scale contextual information and enhancing the target-background contrast features are crucial for improving the robustness of infrared small target recognition. Summary of the Invention

[0004] To address the aforementioned technical problems in existing technologies, this invention proposes a dual-branch infrared small target recognition method based on selective state space. By introducing prior knowledge of target edges and combining dynamic feature filtering and cross-level boundary fusion mechanisms, the detection accuracy and contour integrity of small targets in complex backgrounds are significantly improved. The specific technical solution is as follows:

[0005] A dual-branch infrared small target recognition method based on selective state space includes:

[0006] Step 1: Collect an infrared small target dataset and divide the dataset into a training set, a validation set, and a test set;

[0007] Step 2: Input the infrared small target dataset into the backbone network ResNet-50 for feature extraction, and then input it into a dual-branch structure, which includes the Region of Interest Prediction (RPN) module and the Attention Aggregation (KAAM) module.

[0008] Step 3: After feature extraction by the backbone network ResNet-50, the RPN module generates bounding boxes for candidate regions (ROIs) of potential predicted objects. The background image outside the bounding boxes is cleared to obtain the location information map of the predicted object. Then, it is input into MAMBA BLOCK to obtain the location information map of small targets.

[0009] Step 4: Input the multi-scale features extracted by ResNet50 into the KAAM module to perform differential enhancement on the information of each channel, generate a three-channel feature map, and then fuse it with the small target location information map obtained in Step 3 at the feature level to construct an enhanced feature map that combines semantic representation and salient location information.

[0010] Step 5: Input the enhanced feature map from Step 4 into the target discrimination network. The target discrimination network uses candidate regions as a basis to perform deep feature extraction on the enhanced feature map through multi-layer residual blocks, introduces a sequence modeling structure to construct the contextual relationship between regions, and inputs the extracted features pairwise into the MRFFM module for cross-scale residual feature fusion to obtain the final fused feature map. Then, after further feature compression and classification head mapping, the binary classification result is output to determine whether each candidate region is a real small target, thereby achieving target screening.

[0011] Step 6: Input the final fused feature map from Step 5 and the predicted object location information map output from Step 3 together for fusion, and generate a fused feature map after semantic feature reorganization, guiding the final segmentation head to generate a more discriminative target mask map.

[0012] Further, step 1 specifically involves collecting two datasets: dataset 1 and dataset 2, where dataset 1 contains infrared small target images and dataset 2 contains tooth images. Images are selected from dataset 1 and dataset 2 respectively as training set, validation set, and test set.

[0013] Furthermore, in step 2, the Resnet-50 network extracts coarse features of small infrared targets and then inputs them into the KAAM module for important feature enhancement and local edge weakening.

[0014] Furthermore, the KAAM module includes a channel attention modeling branch and a spatial attention modeling branch. First, the extracted coarse features of small infrared targets are input into the channel attention modeling branch. X Global average pooling and max pooling operations are performed respectively to obtain two channel description vectors. and Then, the channel description vectors are input into the Efficient KAN Block, and a nonlinear feature mapping is performed using a learnable spline interpolation function. The outputs of the two branches are added and fused along the channel dimension, and then a channel attention map is generated by passing it through a Sigmoid activation function. A c Finally, the channel attention map is multiplied channel by channel with the original input features to obtain the channel-enhanced intermediate feature map. X c After that, X c The input to the spatial attention modeling branch first performs max pooling and average pooling along the channel dimension to generate two spatial feature maps. The two maps are then concatenated along the channel dimension and fed into a 7×7 convolutional layer, followed by sigmoid activation to generate the spatial attention map. Ultimately, the spatial attention map is applied to X cObtain the output feature map of the KAAM module X out .

[0015] Furthermore, in step 3, the Region of Interest (RPN) module extracts multi-scale feature maps from the output of the backbone network ResNet-50, extracts spatial location information of potential target regions, and generates a set of candidate ROIs by combining anchor box mechanism or dense sliding window strategy. Each candidate ROI is accompanied by location coordinates and foreground probability score to indicate the likelihood that it contains a target.

[0016] Furthermore, in step 3, the MAMBA BLOCK as a whole comprises two stacked sub-modules. Each sub-module consists of a batch normalization layer, an SS2D spatial selective modeling module, and a learnable activation function package, and a residual connection is introduced between the two sub-modules. The predicted object location information map is input into the batch normalization layer. E The mean and variance of the standardized input features are then fed into the SS2D spatial selective modeling module. This module selectively models information along the spatial dimension H×W, uncovering latent spatial patterns. The output intermediate features are then fed into a learnable activation function package for nonlinear transformation. The activation mechanism of this learnable activation function package is based on a weighted combination of multiple basic activation functions. The activated output results are then processed... F The input and output features are fused through short-circuit connections. The fused features then pass through a batch normalization layer, an SS2D spatial selective modeling module, and a learnable activation function package to form the final output features. F 2. This refers to the location information map of small targets.

[0017] Furthermore, in step 5, the specific process of deep feature extraction by the target discrimination branch includes:

[0018] Step 5.1.1: Based on the enhanced feature map and candidate bounding box set, extract the feature regions by aligning the region of interest using RoIAlign. f i ;

[0019] Step 5.1.2: Extract each feature region f i Several layers of residual blocks are fed in to extract deep semantic features. ;

[0020] Step 5.1.3: Add all RoI features Treat it as a sequence To perform regional-level context modeling.

[0021] Furthermore, the specific steps for cross-scale residual feature fusion in the MRFFM multi-scale residual feature fusion module include:

[0022] Step 5.2.1: Input two feature maps of different scales and First, sum the two elements one by one to obtain the fusion basic features. F sum ;

[0023] Step 5.2.2: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] F sum The input is fed into two independent gate units, Gate1 and Gate2, to extract attention features from different directions or channels, resulting in two attention vectors: A1 = Gate1( Fsum ),A2=Gate2( Fsum );

[0024] Step 5.2.3: Input attention vectors A1 and A2 into the SS2D module for orientation modeling, and normalize them using the Softmax operation to generate attention weight maps W1=Softmax(SS2D(A1)), W2=Softmax(SS2D(A2));

[0025] Step 5.2.4: Compare the obtained attention weight map with the original input. F1 , F2 Multiplication achieves weighted summation;

[0026] Step 5.2.5: Finally, calculate the weighted feature map. F1 ， and F2 ， Element-wise summation is performed to achieve residual feature fusion, resulting in the final fused feature map. F MRFFM This completes the fusion of multi-scale residual features.

[0027] The present invention has the following advantages:

[0028] (1) Based on the backbone feature extraction network, this invention introduces the Kolmogorov-Arnold attention module, which adaptively strengthens the response region of small targets through the channel-level differential feature enhancement mechanism, effectively improving the detection accuracy of infrared small targets.

[0029] (2) The present invention designs an MRFFM module, which uses a selective state space structure to perform contextual modeling of multi-scale fusion features, thereby enhancing the model’s ability to perceive long-distance dependencies between targets.

[0030] (3) This invention proposes a dual-branch feature extraction architecture, which captures local detail information and global semantic information respectively, and uses information complementarity mechanism to achieve accurate recognition of small targets in complex backgrounds.

[0031] (4) The present invention constructs a learnable activation residual block and introduces a combination structure of multiple activation functions to improve the nonlinear expression capability of the model and effectively enhance the feature modeling capability of weak infrared targets.

[0032] (5) The present invention introduces a contrast constraint strategy during the training process, which improves the discrimination ability and robustness of the model by constructing feature distance constraints between the target and the background. Attached Figure Description

[0033] Figure 1 This is a diagram of a dual-branch infrared small target recognition network based on selective state space according to an embodiment of the present invention;

[0034] Figure 2 This is a schematic diagram of the KAAM module according to an embodiment of the present invention;

[0035] Figure 3 This is a schematic diagram of the MRFFM multi-scale residual feature fusion module according to an embodiment of the present invention;

[0036] Figure 4 This is a schematic diagram of a learnable activation residual block according to an embodiment of the present invention. Detailed Implementation

[0037] To make the objectives, technical solutions, and technical effects of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0038] like Figure 1 As shown, this embodiment discloses a dual-branch infrared small target recognition method based on selective state space, including the following steps:

[0039] Step 1: Collect an infrared small target dataset and divide the dataset into a training set, a validation set, and a test set.

[0040] Specifically, two datasets are collected: D = {dataset 1, dataset 2}, where dataset 1 contains 1000 infrared small target images and dataset 2 contains 1329 tooth images. 700 and 1000 tooth images are selected from dataset 1 and dataset 2, respectively, as the training set, 200 and 229 as the validation set, and the rest as the test set.

[0041] Step 2: The infrared small target dataset is input into the backbone network ResNet-50 for feature extraction, followed by a dual-branch structure including a Region of Interest (RPN) module and an Attention Aggregation (KAAM) module. The ResNet-50 backbone network extracts coarse features of the infrared small targets. X The data is then fed into the KAAM module to add important features and weaken local edges, further enhancing the model's ability to focus on and extract important features.

[0042] Step 3: Input the features extracted by the backbone network ResNet-50 into the Region of Interest (RPN) prediction module. The RPN module generates bounding boxes for candidate regions of potential predicted objects and removes the background image except for the bounding boxes to obtain a predicted object location information map E. This map is then input into the MAMBA BLOCK to obtain the location information map of small targets, improving the network's ability to model long-distance spatial and inter-channel dependencies within candidate regions and constructing high-level semantic features for subsequent classification.

[0043] Specifically, the Region of Interest (RPI) prediction module is used to quickly and efficiently define preliminary bounding boxes for regions in infrared images that may contain small targets. This module extracts spatial location information of potential target regions based on multi-scale feature maps and, combined with anchor-box mechanisms or dense sliding window strategies, generates a set of candidate RPIs. Each candidate RPI is accompanied by location coordinates and a foreground probability score to indicate its likelihood of containing a target; and background is cleared. This step not only significantly reduces the search space and the computational complexity of subsequent processing but also improves the recall and accuracy of small target detection.

[0044] After the RPN processes the candidate regions of interest, the MAMBA BLOCK is applied. The MAMBA BLOCK is used to further perform deep feature modeling and nonlinear transformations on the candidate regions of interest to enhance the separability of small targets. Its structure is as follows: Figure 4 As shown, the MAMBA BLOCK consists of two stacked sub-modules. Each sub-module comprises a Batch Normalization layer, an SS2D spatial selective modeling module, and a learnable activation function package. Residual connections are introduced between the two sub-modules to improve the stability and generalization ability of the network training. First, the input features... After passing through the Batch Normalization layer, the mean and variance of the input features are standardized to accelerate convergence and prevent gradient explosion. The normalized features are then fed into the SS2D module, which selectively models information in the spatial dimension H×W, uncovering latent spatial patterns. The output intermediate features are then fed into a learnable activation function package for nonlinear transformation. This activation mechanism does not use a single function but rather a weighted combination of multiple basic activation functions. The activated output is then processed... F The input and output features are fused through a short-circuit connection. The fused features then pass through a BN layer, an SS2D module, and an activation function package to form the final output. The MAMBA BLOCK module outputs the features. F 2 refers to the small target location information map. This feature contains higher-order, non-linear, and richer contextual semantic information, which helps the subsequent detection and recognition branches to accurately identify infrared small targets.

[0045] After the above steps, the predicted object location information map E obtained from RPN is input into the MAMBA BLOCK module, and higher-order feature enhancement is achieved through the above processing.

[0046] The relevant calculation formulas for this module are as follows:

[0047] ,

[0048] ,

[0049] ,limited ∈ Learnable function package;

[0050] ,

[0051] ,

[0052] ,

[0053] ,

[0054] ,

[0055] ,

[0056] ,

[0057] ,

[0058] In the formula, ax In a It is learnable. This represents the output features after passing through a learnable activation function package; This represents the i-th activation function; this package uses a weighted combination of multiple basic activation functions to improve nonlinear modeling capabilities; represents the learnable weight coefficient of the i-th activation function, reflecting the contribution of the activation function in the combination, and satisfies the normalization constraint; BN(·): represents the batch normalization operation, which standardizes the feature distribution; SS2D(·): represents the Selective Spatial 2D module, which is an operation designed to highlight spatially salient regions and further enhance the information of small targets; Xd Along direction d Expanded sequence Conv1Dd represents the direction. d 1D convolution on; A d ,B d ,C d :direction d The corresponding state-space kernel parameters; w d :direction d Learnable fusion weights; L d : The sequence length in this direction.

[0059] Step 4: Input the multi-scale features extracted by ResNet50 in Step 2 into the KAAM module to perform differential enhancement on the information of each channel and generate a three-channel feature map. Then, fuse it with the small target location information map output by Mamba Block in Step 3 at the feature level to construct an enhanced feature map that combines semantic representation and salient location information, providing a unified and robust perceptual foundation for subsequent segmentation and discrimination.

[0060] like Figure 2 As shown, the KAAM module is composed of a channel attention modeling branch and a spatial attention modeling branch, which can effectively enhance the response features of the target structure in high-density artifact regions and improve the feature extraction capability of boundary regions.

[0061] The implementation of this module is as follows: first, the effective features... The input is fed into the channel attention modeling branch, where spatial statistics are calculated along the channel dimension of the input feature map. Global average pooling and max pooling operations are then performed to obtain two channel description vectors: mean feature vector and mean feature vector. Peak characteristics .

[0062] The channel description vectors are then input into the Efficient KAN Block, where a learnable spline interpolation function is used for nonlinear feature mapping to enhance the ability to model inter-channel relationships. The outputs of the two branches are added together along the channel dimension and then fused using a sigmoid activation function to generate a channel attention map. A c Finally, the channel attention weights are multiplied channel by channel with the original input features to obtain the channel-enhanced intermediate feature map. X c .

[0063] Then X c The input is fed into the spatial attention modeling branch. First, max pooling and average pooling are performed along the channel dimension to generate two spatial feature maps. The two maps are then concatenated along the channel dimension and fed into a 7×7 convolutional layer to extract spatial context information. Finally, a sigmoid activation is applied to generate the spatial attention map. Finally, the spatial attention map is applied to the element-wise multiplication circle. X c Highlighting key spatial regions, the output feature map of the KAAM module is obtained. X out .

[0064] The relevant calculation formulas for this module are as follows:

[0065] ,

[0066] ,

[0067] ,

[0068] ,

[0069]

[0070] ,

[0071] In the formula This indicates the input feature map X Average pooling is performed in the spatial dimension to calculate the average value across all positions within c channels; while `c` represents the maximum value in the `c`-th channel, used to extract the most significant response within that channel. `EKAN(⋅)` represents the Efficient KAN Block, used to perform a non-linear mapping on the pooled vectors to enhance their expressive power; `add(⋅)` represents element-wise addition of the two mapped vectors. Channel attention weights are generated using the Sigmoid activation function. A cThen, a channel-by-channel multiplication operation is performed with the original input feature map X to obtain the channel-enhanced features. X c Subsequently, in spatial attention modeling, and They represent spatial locations respectively (i,j) The average and maximum values ​​of all channels; Concat(⋅) This indicates that the two graphs are concatenated along the channel dimension; Conv7×7 indicates that a convolution operation with a kernel size of 7×7 is used to model spatial context dependencies. Represented by the Sigmoid activation function, it compresses the output values ​​to the [0, 1] interval, representing the importance weight of each channel. X out for A s and X c Element-wise multiplication is performed to obtain the final output feature map after joint enhancement of channel and spatial attention.

[0072] The KAAM module utilizes dual-pooling feature fusion and channel-space joint attention mechanisms to refine and enhance the coarse features extracted by the backbone network in multiple dimensions. After the above processing, the obtained... X out The coarse features extracted from the backbone network were enhanced by incorporating information from channel and spatial tooth features, thereby improving the model's ability to perceive detailed tooth features.

[0073] Step 5: Input the enhanced feature map fused in Step 4 into the target discrimination network. This network uses candidate box regions as a basis to perform deep feature extraction through multi-layer residual blocks, introduces a sequence modeling structure to construct contextual relationships between regions, and inputs the extracted features pairwise into the MRFFM multi-scale residual feature fusion module for cross-scale residual feature fusion. After further feature compression and classification head mapping, the binary classification result is output to determine whether each candidate region is a real small target, thereby achieving target selection with saliency enhancement and background suppression.

[0074] Specifically, the deep feature extraction process performed by the target discrimination network includes:

[0075] Step 5.1.1: Enhance the feature map as follows X out1 ∈R C×H×W The candidate box region set is as follows: R= { r 1 ,r 2 ,…,r K For each candidate boxr i Feature regions are extracted using RoIAlign:

[0076] ,

[0077] Step 5.1.2: Multi-layer residual block processing, processing each f i Several layers of residual blocks are fed in to extract deep semantic features:

[0078] ,

[0079] The residual block structure is as follows:

[0080] ,

[0081] In this formula, F represents two convolutional layers and normalization (BN-ReLU-Conv), and W={W1,W2} represents the weights of the two convolutional layers.

[0082] Step 5.1.3: Sequence modeling, treating all RoI features as a sequence. To perform region-level context modeling, the specific calculation formula is as follows:

[0083] ,

[0084] In the formula This represents the input features from the i-th candidate region after processing the residual block. The features representing the j-th candidate region, This represents the learnable parameter matrix, which maps input features to a query vector. This represents mapping the input features to a key vector. This represents mapping the input features to a value vector. Let d represent the attention weight of j with respect to i, and d represent the scaling factor of the feature dimension. This can be expressed as a softmax operation on the attention scores for all j=1,…,K, such that… .

[0085] like Figure 3 As shown, the specific steps for cross-scale residual feature fusion in the MRFFM multi-scale residual feature fusion module include:

[0086] Step 5.2.1: Input two feature maps of different scales and First, sum the two elements one by one to obtain the fusion basic features. F sum ;

[0087] Step 5.2.2: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation F sum The input is fed into two independent gate units, Gate1 and Gate2, to extract attention features from different directions or channels, resulting in two attention vectors: A1 = Gate1( F sum ),A2=Gate2( F sum );

[0088] Step 5.2.3: Input attention vectors A1 and A2 into the SS2D module for orientation modeling, and normalize them using the Softmax operation to generate attention weight maps W1=Softmax(SS2D(A1)), W2=Softmax(SS2D(A2));

[0089] Step 5.2.4: Compare the obtained weight map with the original input. F1 , F2 Multiplication achieves weighted summation. F1′ =W1⊙ F1 , F2′ =W2⊙ F2 ;

[0090] Step 5.2.5: Finally, calculate the weighted feature map. F1′ and F2′ Element-wise summation is performed to achieve residual feature fusion, resulting in the final fused feature map F. MRFFM This completes the fusion of multi-scale residual features.

[0091] Thus, after the above steps, the fused feature map completes explicit relation modeling and attention-guided enhancement in the MRFFM module, laying a more solid semantic foundation for subsequent target discrimination and fine classification.

[0092] Step 6: Based on the final fused feature map obtained in Step 5 F MRFFM The predicted object location information map E output from step 3 is input and fused together with it. After semantic feature reorganization, a fused feature map is generated to guide the final segmentation head in generating a more discriminative target mask map. The expression is as follows:

[0093] ,

[0094] in, Used to semantically reorganize fused features and enhance contextual consistency, SegHead represents a standard convolution classifier used to generate the final mask prediction result.

[0095] To demonstrate the effectiveness of this method, the datasets used in this experiment are from the publicly available datasets NUAA-SIRST and WideIRSTD. The labeled datasets provided by the competition contain 1200 and 400 images respectively, which, after data augmentation, result in a total of 4800 and 1600 images. The datasets are divided into training, validation, and test sets. The training set is used to optimize the parameters of the neural network, while the validation set is used to evaluate the model's applicability to data outside the training set and to determine the required number of iterations. After model training, the test set is used to evaluate quality metrics. The dataset segmentation is based on typical values ​​commonly used in the task; specifically, images are segmented into training, validation, and test sets at ratios of 70%, 10%, and 20%; the input dataset size is 256×256. The test results are compared with mainstream tooth segmentation algorithms in the existing technology. The proposed model framework is implemented in PyTorch, trained using an 8GB GPU, and the Adam optimization algorithm is used to optimize the overall parameters with a learning rate set to 0.01e-4. The test results for the public datasets of NUAA-SIRST and WideIRSTD are shown in Table 1:

[0096] Table 1:

[0097]

[0098] Table 1 shows the test results of this invention and five other methods on the public datasets NUAA-SIRST and WideIRSTD for tooth segmentation. The evaluation metrics used in the table are all commonly used in infrared small target detection tasks, and the results are positively correlated with the metrics.

[0099] In summary, this invention proposes a dual-branch infrared small target recognition method based on selective state space, relating to the field of infrared image processing. This method constructs a (SSS-DB IRSTN) selective state space dual-branch infrared small target recognition network for training, mainly including a backbone network, a KAAM module, a Mamba feature fusion and relation modeling module, and learnable activation residual blocks. The optimal model obtained after training is used to obtain the infrared small target results. This method deeply integrates multi-scale features and spatial dynamic cues to improve the accuracy of infrared small target detection in low signal-to-noise ratio environments. The dual-branch architecture extracts local details and global contextual information separately, and the selective state space module is introduced to dynamically model the spatiotemporal dependency between the target and the background, effectively overcoming the limitations of traditional algorithms in recognizing complex background interference and target size variations. Furthermore, this invention designs an optimization strategy based on contrast constraints during training to further enhance the network's accuracy in identifying and segmenting small targets, achieving highly robust small target recognition performance.

[0100] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any way. Although the implementation process of the present invention has been described in detail above, those skilled in the art can still modify the technical solutions described in the foregoing examples or make equivalent substitutions for some of the technical features. All modifications and equivalent substitutions made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A dual-branch infrared small target recognition method based on selective state space, characterized in that, include: Step 1: Collect an infrared small target dataset and divide the dataset into a training set, a validation set, and a test set; Step 2: Input the infrared small target dataset into the backbone network ResNet-50 for feature extraction, and then input it into a dual-branch structure, which includes the Region of Interest Prediction (RPN) module and the Attention Aggregation (KAAM) module. The backbone network Resnet-50 extracts coarse features of small infrared targets and then inputs them into the KAAM module to add important features and weaken local edges. The KAAM module includes a channel attention modeling branch and a spatial attention modeling branch. First, the extracted coarse features of small infrared targets are input into the channel attention modeling branch. X Global average pooling and max pooling operations are performed respectively to obtain two channel description vectors. and Then, the channel description vectors are input into the Efficient KAN Block, and a nonlinear feature mapping is performed using a learnable spline interpolation function. The outputs of the two branches are added and fused along the channel dimension, and then a channel attention map is generated by passing it through a Sigmoid activation function. A c Finally, the channel attention map is multiplied channel by channel with the original input features to obtain the channel-enhanced intermediate feature map. X c After that, X c The input to the spatial attention modeling branch first performs max pooling and average pooling along the channel dimension to generate two spatial feature maps. The two maps are then concatenated along the channel dimension and fed into a 7×7 convolutional layer, followed by sigmoid activation to generate the spatial attention map. Ultimately, the spatial attention map is applied to X c Obtain the output feature map of the KAAM module X out ; Step 3: After feature extraction by the backbone network ResNet-50, the RPN module generates bounding boxes for candidate regions (ROIs) of potential predicted objects and removes the background image except for the bounding boxes to obtain the predicted object location information map. Then, it is input into the MAMBA BLOCK module to obtain the small target location information map. Step 4: Input the multi-scale features extracted by ResNet-50 into the KAAM module to perform differential enhancement on the information of each channel, generate a three-channel feature map, and then fuse it with the small target location information map obtained in Step 3 at the feature level to construct an enhanced feature map that combines semantic representation and location information. Step 5: Input the enhanced feature map from Step 4 into the target discrimination network. The target discrimination network uses candidate regions as a basis to perform deep feature extraction on the enhanced feature map through multi-layer residual blocks, introduces a sequence modeling structure to construct the contextual relationship between regions, and inputs the extracted features pairwise into the MRFFM module for cross-scale residual feature fusion to obtain the final fused feature map. Then, after further feature compression and classification head mapping, the binary classification result is output to determine whether each candidate region is a real small target, thereby achieving target screening. Step 6: Input the final fused feature map from Step 5 and the predicted object location information map output from Step 3 together for fusion, and generate a fused feature map after semantic feature reorganization, which guides the final segmentation head to generate a more discriminative target mask map.

2. The dual-branch infrared small target recognition method based on selective state space according to claim 1, characterized in that, Step 1 specifically involves collecting two datasets: dataset 1 and dataset 2. Dataset 1 contains infrared small target images, and dataset 2 contains tooth images. Images are selected from dataset 1 and dataset 2 respectively as the training set, validation set, and test set.

3. The dual-branch infrared small target recognition method based on selective state space according to claim 1, characterized in that, In step 3, the Region of Interest (RPN) module extracts multi-scale feature maps from the output of the backbone network ResNet-50, extracts spatial location information of potential target regions, and generates a set of candidate ROIs by combining anchor box mechanism or dense sliding window strategy. Each candidate ROI is accompanied by location coordinates and foreground probability score to indicate the likelihood that it contains a target.

4. The dual-branch infrared small target recognition method based on selective state space according to claim 1, characterized in that, In step 3, the MAMBA BLOCK as a whole comprises two stacked sub-modules. Each sub-module consists of a batch normalization layer, an SS2D spatial selective modeling module, and a learnable activation function package, with residual connections introduced between the two sub-modules. The predicted object location information map is input into the batch normalization layer. E The mean and variance of the standardized input features are then fed into the SS2D spatial selective modeling module. This module selectively models information along the spatial dimension H×W, uncovering latent spatial patterns. The output intermediate features are then fed into a learnable activation function package for nonlinear transformation. The activation mechanism of this learnable activation function package is based on a weighted combination of multiple basic activation functions. The activated output results are then processed... F The input and output features are fused through short-circuit connections. The fused features then pass through a batch normalization layer, an SS2D spatial selective modeling module, and a learnable activation function package to form the final output features. F 2. This refers to the location information map of small targets.

5. The dual-branch infrared small target recognition method based on selective state space according to claim 4, characterized in that, In step 5, the specific process of deep feature extraction by the target discrimination branch includes: Step 5.1.1: Based on the enhanced feature map and candidate bounding box set, extract the feature regions by aligning the region of interest using RoIAlign. f i ; Step 5.1.2: Extract each feature region f i Several layers of residual blocks are fed in to extract deep semantic features. ; Step 5.1.3: Add all RoI features Consider it as a sequence { f 1 ′,f 2 ′,…,f K ′ }, to perform regional-level context modeling.

6. The dual-branch infrared small target recognition method based on selective state space according to claim 5, characterized in that, The specific steps for cross-scale residual feature fusion in the MRFFM multi-scale residual feature fusion module include: Step 5.2.1: Input two feature maps of different scales and First, sum the two elements one by one to obtain the fusion basic features. F sum ; Step 5.2.2: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] F sum The input is fed into two independent gate units, Gate1 and Gate2, to extract attention features from different directions or channels, resulting in two attention vectors: A1=Gate1(Fsum), A2=Gate2(Fsum); Step 5.2.3: Input attention vectors A1 and A2 into the SS2D module for orientation modeling, and normalize them using the Softmax operation to generate attention weight maps W1=Softmax(SS2D(A1)), W2=Softmax(SS2D(A2)); Step 5.2.4: Compare the obtained attention weight map with the original input. F1 , F2 Multiplication achieves weighted summation; Step 5.2.5: Finally, calculate the weighted feature map. F1 ， and F2 ， Element-wise summation is performed to achieve residual feature fusion, resulting in the final fused feature map. F MRFFM This completes the fusion of multi-scale residual features.

Images