Infrared small target detection method fusing local prior and multi-scale global background

Abstract

The invention discloses an infrared small target detection method fusing local prior and a multi-scale global background, and the method comprises the steps: firstly obtaining image data containing an infrared image and a mask label corresponding to the infrared image, and carrying out the preprocessing; secondly, a target detection model of an encoder-decoder architecture is constructed, an encoder comprises a local detail prior mining branch and a multi-scale global background perception branch which are parallel, step-by-step feature extraction is performed on the preprocessed image data, and a decoder comprises a progressive feature fusion decoding branch; and inputting the features of each level of the encoder double branches into decoder branches for decoding step by step to obtain a detection result. And finally, a weighted depth supervision mechanism is introduced in training, auxiliary prediction output is set in a plurality of decoding layers, and weighting loss is calculated. According to the method, the problems of insufficient local detail modeling, insufficient multi-scale global background perception of Mamba, difficulty in global and local feature fusion and the like in the existing method are solved, and the detection precision of the infrared small target is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of infrared small target detection, and in particular relates to an infrared small target detection method that integrates local priors and multi-scale global background. Background Art

[0002] Infrared imaging equipment offers advantages such as all-weather operation, strong anti-interference resistance, and the ability to detect camouflaged targets. However, in complex backgrounds and long-range imaging conditions, small targets in infrared images are often extremely small, have weak textures, and low contrast, easily being drowned out by background noise, posing a significant challenge to automatic detection. Because small targets occupy only a few pixels, traditional bounding box regression-based methods are less robust and prone to missed detections or false positives. Therefore, existing research generally models infrared small target detection as an image segmentation problem.

[0003] Existing infrared small target detection methods primarily include traditional methods, methods based on convolutional networks (CNNs), methods based on Transformers, and a small number of Mamba-based state-space modeling methods. Traditional methods rely on manually designed features and prior rules, such as local contrast, filter response, or sparse modeling. These methods are generally sensitive to parameters and have poor generalization capabilities, making them incapable of adapting to complex and changing real-world backgrounds. CNN-based methods automatically learn feature representations through end-to-end training. However, due to their limited receptive fields, they struggle to model large-scale contextual information, resulting in unstable target-level detection performance and prone to false alarms or missed detections. The Transformer introduces a global self-attention mechanism, enhancing its ability to model long-range dependencies. However, this mechanism suffers from high computational overhead, limited real-time performance, and the weak signal and low-texture characteristics of infrared images, which can easily degrade the attention mechanism's expressive power. Mamba, a recently proposed state-space modeling framework, combines linear computational complexity with strong global modeling capabilities, showing promising results in natural language processing and vision tasks. However, existing work has not yet effectively addressed the core challenges of this task: First, existing methods lack the ability to model prior information on local details, making it difficult to capture pixel contrast between the target and the background and gradient changes at the target's edges, resulting in weak feature responses for small targets. Second, Mamba's fixed-scale image segmentation method limits its multi-scale background perception capabilities, making it difficult to adapt to the distribution characteristics of targets of different scales in complex backgrounds. Finally, there is a semantic deviation between the global semantic information and local detail features modeled by Mamba, making them difficult to align and fuse, affecting the precise positioning of targets and the restoration of their boundaries.

[0004] Therefore, how to take advantage of Mamba's linear computational complexity to design an infrared small target detection method that has the ability to mine local prior information and perceive multi-scale global background context, and can effectively integrate global and local features, has become an important technical problem that needs to be solved urgently. Summary of the Invention

[0005] To address the above problems, the present invention aims to propose a method for infrared small target detection that integrates local priors and multi-scale global background information, addressing the problems of insufficient local prior information capture, insufficient multi-scale global background context perception, and difficulty fusing global and local features when applying Mamba to small infrared targets. To achieve the above objectives, the present invention specifically adopts the following technical solutions:

[0006] The infrared small target detection method that integrates local prior and multi-scale global background includes the following steps:

[0007] Step 1: Obtain PNG format image data containing the infrared image and its corresponding binary mask label, and perform preprocessing, including random rotation, Gaussian processing, and uniform cropping operations on the original data.

[0008] Step 2. Design a neural network model for the infrared small target detection method. The model is an encoder-decoder architecture, in which the encoder contains parallel local detail prior mining branches and multi-scale global background perception branches, and the decoder contains a progressive feature fusion decoding branch. The encoder is used to extract features from the preprocessed image data step by step, and the features of each level of the encoder's two branches are input into the decoder branch, and the detection results are obtained by decoding step by step.

[0009] Step 2.1: Construct a local detail prior mining branch, which consists of Layer details capture convolutional blocks ( ) are connected in series to extract prior information such as target-background pixel contrast difference and target edge gradient change. The input of this branch is infrared image. layer Then a local feature map will be output .

[0010] Step 2.2: Construct a multi-scale global background perception branch, which consists of Layer multi-scale Mamba block ( ) are connected in series to realize the perception and fusion of background semantic information at different scales. The input of this branch is infrared image, which is input into the series after the patch expansion operation (PE). , No. layer Afterwards, a global feature map will be output through the patch merging operation (PM) .

[0011] Step 2.3: Construct a progressive fusion decoding branch, which consists of Layer feature fusion decoding module ( ) are connected in series to gradually restore the spatial resolution and generate the target mask; layer Receive and fuse a feature map from the multi-scale global background perception branch step by step , a feature map of a local detail prior mining branch and a lower layer fusion decoding output feature map , among which layer The lower layer fusion decoding output to be received is by and After processing through concatenation, basic convolution block, Squeeze-and-Excitation (SE) attention mechanism and another basic convolution block, a fused decoding feature map of this layer is output. .

[0012] Step 3: A weighted deep supervision mechanism is introduced during training. This involves setting auxiliary prediction outputs at multiple decoding layers and calculating weighted losses for each, forming a multi-level supervision signal to accelerate gradient backpropagation. Loss calculation uses a joint loss function consisting of the soft intersection over union (SoftIoU) loss and the binary cross entropy loss (BCE). The SoftIoU loss optimizes the outline overlap and segmentation consistency of the target area, while the BCE improves the pixel-level classification accuracy of the target and background.

[0013] Step 4: Evaluate and apply the trained model.

[0014] A further technical solution is that the specific solution of the local detail prior mining branch in step 2 is as follows: this branch is mainly composed of four layers of detail-capturing convolution blocks connected in series. The detail-capturing convolution block first expands the feature dimension through a basic convolution block, and then the feature input is paralleled by an original convolution kernel and four differential convolution kernels. The four differential convolution kernels are center differential convolution (CDC), angle differential convolution (ADC), horizontal differential convolution (HDC) and vertical differential convolution (VDC). CDC and ADC are used to extract the pixel difference prior between the target and the background, and HDC and VDC are combined to extract the target edge gradient change prior. After the convolution results are added and fused, they are processed by a batch normalization and ReLU activation function to output a local feature map. The basic convolution block is composed of a cascade of convolution layers, batch normalization layers and ReLU activation function layers.

[0015] A further technical solution is to implement the multi-scale global context perception branch in step 3 as follows: This branch consists of four layers of multi-scale Mamba blocks connected in series. A depthwise separable convolution block is first used within the multi-scale Mamba block for channel expansion. A classic visual state space block (VSS block) is then used to model long-range dependencies through the state space mechanism of the two-dimensional selective scan operator (SS2D). State recursion is performed on the input feature map in both the horizontal and vertical directions. Global state perception is then achieved through a linear layer and a GELU activation function. After global state perception is achieved, to improve the model's adaptability to objects and background structures of varying scales, the VSS block is followed by three parallel depthwise separable convolution blocks of varying scales, with progressively larger kernel sizes. The three parallel output feature maps are fused by addition and finally integrated using another linear layer. This output integrates a high-level representation of multi-scale global context information.

[0016] A further technical solution is that the progressive feature fusion decoding branch scheme in step 4 is as follows: the branch is mainly composed of four layers of feature fusion decoding modules in series, which are used to decode the feature map from the global branch in the decoding stage. , feature map of local branch And the fusion decoding output feature map of the lower layer Perform step-by-step fusion to improve the consistency of feature representation and the accuracy of target positioning. layer, will 、 and After splicing, the basic convolution block is used to reduce the channel to obtain the splicing features , and then use the spatial activation strategy and channel filtering strategy to alleviate the differences in scale and semantic level between global semantic features and local detail features. Finally, the global features of the spatial activation strategy and the local features of the channel filtering strategy are added after the basic convolution block processing dimensions, and are processed by layer normalization and ReLU activation function to obtain , and then passed to the upper layer decoding module to achieve step-by-step feature integration from coarse to fine.

[0017] A further technical solution is: the specific scheme of the spatial activation strategy and channel filtering strategy is as follows: the spatial activation strategy is used to stitch the features Average pooling and maximum pooling operations are applied to the channel dimension respectively. After splicing, the spatial attention map is extracted by convolution and SigMoid activation function, and then combined with the global feature map. Multiply pixel by pixel. At the same time, the channel filtering strategy is to Perform global average pooling and maximum pooling operations, input to the shared multi-layer perceptron and add them, then process them through SigMoid to generate channel attention weights, and compare them with local feature maps one by one. Channel multiplication.

[0018] A further technical solution is that in step 5, the specific solution of the weighted depth supervision mechanism is to set up auxiliary output channels at each level of the decoder to supervise and guide the intermediate feature maps of different semantic depths. The prediction head, consisting of convolution and sigmoid activation functions, generates a small object probability map of a size that matches the label as an auxiliary output. This auxiliary output is compared with the true label map for loss calculation, using a weighted combination of soft intersection over union (SoftIoU) loss and binary cross entropy loss (BCE) to measure the model's ability to segment objects at different scales.

[0019] The present invention has the following characteristics and beneficial effects:

[0020] The local detail prior mining branch of the present invention has the ability to mine the pixel contrast difference and target edge gradient between the target and the background, effectively enhancing the structural expression of small infrared targets with weak texture and low contrast, solving the problem of insufficient modeling of detail information in existing methods, and improving the detection accuracy and separability of small infrared targets.

[0021] The multi-scale global background context perception branch of the present invention has the ability to model long-range dependencies in the image space dimension. It integrates contextual semantic information through a multi-scale depth-separable convolutional structure, solves the problem that traditional methods are difficult to adapt to complex background changes, and improves the robustness and global modeling ability of the model under multiple background interferences.

[0022] The progressive fusion decoding branch of the present invention adopts channel filtering and spatial activation strategies to effectively guide the collaborative fusion of global and local features during the decoding process, solving the problems of inconsistent feature semantics and blurred boundaries, and improving the target boundary restoration capability and feature expression integrity.

[0023] The overall network structure design of the present invention alleviates the computational complexity problem caused by using Transformer while maintaining detection accuracy, and improves the engineering application feasibility of the method. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 This is the overall architecture diagram of the model proposed in the present invention;

[0025] Figure 2 The local detail prior information and the target information in the multi-scale background described in the present invention;

[0026] Figure 3 The convolutional block structure diagram that can capture the details proposed by the present invention;

[0027] Figure 4 This is the multi-scale Mamba block structure diagram proposed in the present invention;

[0028] Figure 5 This is a structural diagram of the feature fusion decoding module proposed in the present invention;

[0029] Figure 6 A visual comparison of the segmentation results of the present invention and representative methods. DETAILED DESCRIPTION

[0030] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. The following examples will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, unless conflicting, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0031] like Figure 1 As shown, the present invention discloses an infrared small target detection method that integrates local priors and multi-scale global background, including the following steps:

[0032] Step 1: Obtain PNG format image data containing infrared images and their corresponding binary mask labels, and perform random rotation, Gaussian processing, and uniform cropping operations on the original data to enhance sample diversity and model robustness. The enhanced image size is .

[0033] In this embodiment, three public infrared small target datasets are selected: (1) NUAA-SIRST dataset: This dataset is the first public infrared small target dataset, containing 427 images and a total of 480 target instances. The images of this dataset are collected from multiple real natural scenes, making it one of the most widely used datasets to date. (2) NUDT-SIRST dataset: This dataset contains 1327 challenging infrared images, covering five types of scenes: city, field, strong light, ocean and cloud. This dataset has the characteristics of diverse target sizes and complex backgrounds. (3) IRSTD-1k dataset: This dataset contains 1001 real infrared images, covering representative scenes such as ocean, river, mountain and city. The present invention adopts the default division ratio of the dataset proposer, where the training set test and test set division ratio of NUAA-SIRST and NUDT-SIRST is 1:1, while the training set test and test set division ratio of IRSTD-1k is 4:1. The data are uniformly subjected to common data enhancement operations such as random rotation, Gaussian processing and cropping, and finally the final result is obtained. .

[0034] Step 2. Design a neural network model for the infrared small target detection method. The model is an encoder-decoder architecture, in which the encoder contains parallel local detail prior mining branches and multi-scale global background perception branches, and the decoder contains a progressive feature fusion decoding branch. The encoder is used to extract features from the preprocessed image data step by step, and the features of each level of the encoder's two branches are input into the decoder branch, and the detection results are obtained by decoding step by step.

[0035] Step 3: Construct a local detail prior mining branch.

[0036] In this embodiment, the branch consists of four layers of detail-capturing convolutional blocks ( ) are connected in series to extract prior information such as target-background pixel contrast difference and target edge gradient change, thereby enriching the target feature expression and enhancing the recognizability of small infrared targets in weak texture and low contrast backgrounds. The input accepted by this branch is the infrared image. layer Then a local feature map will be output To further illustrate the a priori information, Figure 2 As shown in the figure, the first row (from left to right) represents the original image, the target-background pixel difference prior information, and the target edge gradient prior information. The target in the original image is extremely small and its feature information is weak. By introducing this prior information, the feature representation of small targets can be enriched.

[0037] like Figure 3 As shown, The feature dimension is first expanded internally through a basic convolution block, and then the feature input is paralleled by an original convolution kernel and four differential convolution kernels. The four differential convolution kernels are center differential convolution (CDC), angle differential convolution (ADC), horizontal differential convolution (HDC) and vertical differential convolution (VDC). CDC and ADC are used to extract the pixel difference prior between the target and the background. HDC and VDC are combined to extract the target edge gradient change prior. After the convolution results are added and fused, they are processed by a batch normalization and ReLU activation function to output the local feature map. The basic convolution block is composed of a cascade of convolution layers, batch normalization and ReLU activation function layers. The four differential convolutions are modified from the original convolution kernel, and the original convolution kernel and the convolution kernels of the four differential convolutions are defined as:

[0038]

[0039]

[0040]

[0041]

[0042]

[0043] in represents the original convolution kernel, 、 、 and Represent the horizontal difference convolution kernel, vertical difference convolution kernel, center difference convolution kernel and angle difference convolution kernel respectively. Specifically, let the input features and output features be and ,in , then a complete The feature processing of the module can be defined as:

[0044]

[0045] in represents the activation function, It represents the original convolution combined with four differential convolutions in parallel and then connected in series with a batch normalized difference capture convolution. Represents the basic convolutional block.

[0046] Step 4: Construct a multi-scale global background perception branch.

[0047] In this embodiment, this branch consists of four layers of multi-scale Mamba blocks ( ) are connected in series to realize the perception and fusion of background semantic information at different scales. This branch can effectively improve the modeling ability of complex background interference and the ability to capture multi-scale background context dependencies while maintaining linear computational complexity. The input accepted by this branch is an infrared image. After the image undergoes a patch expansion operation (PE), it is input into the series. , No. layer Afterwards, a global feature map will be output through the patch merging operation (PM) To further illustrate the multi-scale global background information, Figure 2As shown in the figure, the four pictures in the second row from left to right represent the target presentation effect under a small-scale background, the target presentation effect under a medium-scale background, the target presentation effect under a large-scale background, and the target presentation effect under a multi-scale background. It can be seen that as the background scale becomes larger, the target positioning becomes clearer, but more background information interference is also introduced into the receptive field. Therefore, by combining the advantages of multi-scale background information, it can not only retain the detailed information of the target at a small scale and suppress the interference caused by the excessive background, but also obtain more sufficient contextual semantics at a larger scale, thereby alleviating the noise introduced by the single-scale background and achieving accurate and robust positioning of the target.

[0048] like Figure 4 As shown, Internally, a depthwise separable convolution block is first used for channel expansion, followed by a classic visual state space block (VSS Block) to model long-range dependencies through the state space mechanism of the two-dimensional selective scan operator (SS2D). State recursion is performed on the input feature map in the horizontal and vertical directions, thereby achieving global state perception in the spatial dimension. Specifically, let the input feature map be: , the bidirectional state modeling process of SS2D is:

[0049]

[0050]

[0051]

[0052] in, and Respectively represent the status in the row (width) direction and column (height) direction, is the coordinate of the image after segmentation; , are the learnable state transition matrix and input projection matrix, To output the mapping weight matrix, we can further learn the fusion weights of the image in two directions. = 0. After a linear layer and GELU activation function, global state perception is achieved. To improve the model's adaptability to objects and background structures of different scales, this module is further connected in series with three parallel separable convolution blocks of different scales and depths. The convolution kernel sizes are: and , the parallel outputs are fused by addition, and finally another linear layer is used to integrate the final output. Let the input features and output features be ,in , then a complete The module feature processing process can be expressed as:

[0053]

[0054] in represents the activation function, represents a linear layer, Stands for VSS Block. DW represent Depthwise Separable Convolution, PDW Represents parallel and Depthwise Separable Convolution: Ultimately, this module outputs a high-level representation that integrates multi-scale global background context information, effectively improving the distinguishability between foreground objects and complex backgrounds, and providing rich and layered semantic support for subsequent detection and segmentation.

[0055] Step 5: Construct a progressive fusion decoding branch.

[0056] In this embodiment, this branch consists of a four-layer feature fusion decoding module ( ) are connected in series to gradually restore the spatial resolution and generate the target mask. This branch can effectively improve the consistency of feature representation and the accuracy of target positioning. Layer feature fusion decoding module It will receive and fuse a feature map from the multi-scale global background perception branch step by step , a feature map of a local detail prior mining branch and a lower layer fusion decoding output feature map , (the fourth layer The lower layer fusion decoding output to be received is by and After processing through concatenation, basic convolutional blocks, (Squeeze-and-Excitation, SE) attention mechanism and another basic convolutional block, a fused decoding feature map of this layer is output. .

[0057] like Figure 5 As shown, For the feature map from the global branch in the decoding stage , feature map of local branch And the fusion decoding output feature map of the lower layer Perform step-by-step fusion to improve the consistency of feature representation and the accuracy of target positioning. layer, will 、 and Perform splicing to obtain splicing features , then after the basic convolution block performs channel reduction, the spatial activation strategy and channel filtering strategy are used to alleviate the differences in scale and semantic level between the global semantic features and the local detail features. Finally, the global features of the spatial activation strategy and the local features of the channel filtering strategy are added after being processed by the basic convolution block. Finally, they are processed by layer normalization and ReLU activation function and passed to the decoding module of the previous layer to achieve step-by-step feature integration from coarse to fine. Assume that the input features are and , the output features are ,in , then a complete The module feature processing process can be expressed as:

[0058]

[0059]

[0060] in Represents the initial features after splicing, Represents a splicing operation, represents the spatial activation strategy, represents the channel filtering strategy, Represents the basic convolutional block. Through the above dual strategies, the feature fusion module effectively improves the separability of small infrared targets in complex backgrounds, achieves a high degree of complementarity between structural semantics and detail information, and provides more discriminative feature support for the final detection results.

[0061] like Figure 5 As shown, the spatial activation strategy has an important influence on the splicing features. Average pooling and maximum pooling operations are applied to the channel dimension respectively, and the concatenation is performed Convolution extracts the spatial attention map, which is multiplied pixel by pixel with the global feature map, thereby significantly enhancing the spatial response strength of the target area and guiding the model to focus on the small target area more accurately. Perform global average pooling and maximum pooling operations, input to the shared multi-layer perceptron to generate channel attention weights, and multiply them channel by channel with the local feature map to adaptively adjust the response strength of each channel, enhance the semantic channel related to the infrared small target, and suppress redundant background interference. Suppose the input is the splicing feature , spatial activation strategy and channel filtering strategies The detailed processing process can be expressed as follows:

[0062]

[0063]

[0064] in represents the activation function, represents a multilayer perceptron, and Represent average pooling and maximum pooling respectively, represent convolution, Represents a splicing operation

[0065] Step 6: A weighted deep supervision mechanism is introduced during training. This involves setting auxiliary prediction outputs at multiple decoding layers and calculating weighted losses for each layer to form a multi-level supervision signal. This accelerates gradient backpropagation, improves convergence efficiency, and enhances the perception and robustness of intermediate layers for small objects. Loss calculation uses a joint loss function consisting of the soft intersection over union (SoftIoU) loss and the binary cross entropy loss (BCE). The former optimizes the outline overlap and segmentation consistency of the target area, while the latter improves the pixel-level classification accuracy of the target and background.

[0066] like Figure 1 Specifically, each decoding layer is connected to a The prediction head, composed of convolution and sigmoid activation functions, generates a small object probability map of a size matching the label as an auxiliary output. This auxiliary output is compared with the true label map for loss calculation, using a weighted combination of soft intersection over union (SoftIoU) loss and binary cross entropy loss (BCE) to measure the model's ability to segment objects at different scales. The loss calculation process can be expressed as:

[0067]

[0068] in represents the loss value, represents the true mask, , , represents the mask of the output channel of each level of the decoder, where After concatenating the first five layers of auxiliary prediction masks, an additional The final output mask obtained by the prediction head composed of convolution and Sigmoid activation function. All masks will participate in the loss calculation during training, and only This supervision mechanism not only accelerates network convergence and improves training stability, but also enhances the expressiveness and semantic consistency of multi-layer feature maps, significantly improving the model's detection robustness and positioning accuracy for small infrared targets.

[0069] In this example, we used PyTorch 2.0 to build the model on an NVIDIA RTX 4090 GPU with 24GB of video memory. In terms of model initialization, the Xavier initialization method was used to set the weights and biases to promote stable convergence during training. The initial learning rate was set to 0.005 and gradually decreased to , achieving a smoother and more efficient optimization process. Parameter updates use the Adagrad optimization algorithm, which is suitable for cases with sparse gradients. The entire model was trained for 400 batches with a batch size of 8. To ensure fairness in the learning process and objective comparison results, all models were trained from scratch without using pretrained weights.

[0070] Step 7: Evaluate and apply the trained model.

[0071] In order to verify the performance of the present invention, in this embodiment, it is evaluated and compared with multiple advanced methods in the current field, specifically including:

[0072] (1) Traditional methods: including Top-Hat (morphological top-hat transform method), Max-Median (maximum-median filtering method), IPI (infrared block image method) and WSLCM (weighted enhanced local contrast measurement method).

[0073] (2) Deep learning-based methods: including ACM (Asymmetric Context Modulation Network), ALCNet (Attention Local Contrast Network), FC3Net (Feature Compensation and Cross-layer Correlation Mining Network), ISNet (Infrared Shape Network), DNANet (Densely Nested Attention Network), UIUNet (Nested U-Net Network), RDIAN (Receptive Field and Direction Guided Attention Network), AGPCNet (Attention Guided Pyramid Context Network), ABC (Bilinear Correlation Attention Network), MSHNet (Multi-scale Detection Head Network), DMFNet (Dual Encoder Multi-stage Feature Fusion Network), PBT (Progressive Context-aware Transformer), MiM (Nested Mamba Network), ID-UNet (Densely Connected U-Net Network) and MMLNet (Multi-branch Mutually Guided Learning Network).

[0074] In order to comprehensively evaluate the excellent effects of the present invention, the evaluation indicators used include:

[0075] (1) Pixel-level segmentation indicators:

[0076] Intersection over Union (IoU): IoU is a pixel-level evaluation metric.

[0077] Normalized Intersection over Union (nIoU): nIoU is the normalized version of IoU;

[0078] F1 Score: This metric evaluates the missed detection rate and false alarm rate at the pixel level.

[0079] (2) Target-level detection indicators:

[0080] Detection rate ( ): Measures the proportion of correctly detected objects, defined as the number of correctly predicted objects and the total number of targets If the deviation of the center point of the predicted target is less than 3 pixels, the target is considered to be detected correctly.

[0081] False alarm rate ( ): A measure of the number of pixels that were incorrectly predicted to be objects. The total number of pixels predicted as targets among all pixels in the image The ratio of

[0082] The comparison results are shown in Table 1 and Table 2. In all indicators of pixel-level segmentation and target-level detection, the deep learning-based methods are significantly better than the traditional methods. This shows that under complex conditions such as cluttered background or weak targets, traditional methods have limited effects, while deep models can achieve higher accuracy through data-driven learning. Among these methods, the present invention achieved the best overall performance in all indicators and datasets. On the NUAA-SIRST dataset, the present invention achieved the highest IoU thanks to its local prior information mining strategy, which enhances the representation of small targets by modeling the target-background pixel contrast difference and the significance of the target edge. On the NUDT-SIRST dataset, although DMFNet improved the IoU by retaining more shallow features, its performance was poor due to the lack of global context modeling. This is 4.07% lower than the proposed method, highlighting the advantages of the proposed method's multi-scale global background modeling. Even on the challenging IRSTD-1K dataset, the proposed method maintains a good balance between segmentation and detection. This is due to the proposed feature fusion strategy, which effectively integrates features across different levels, narrowing the gap between global and local features and thus improving decoding accuracy.

[0083] Table 1 shows the quantitative comparison results of the proposed method and the existing methods on the NUAA-SIRST dataset. The best results are shown in bold.

[0084]

[0085] Table 2 shows the quantitative comparison results of the proposed method and the existing methods on the NUDT-SISRT and IRSTD-1k datasets. The best results are shown in bold.

[0086]

[0087] To further verify the effectiveness of the present invention, Figure 6 As shown in Figure 1, several typical scenes are randomly selected and compared with the visual detection results of two state-of-the-art methods. The target area is marked with a blue box and magnified when necessary; correct detection, false alarm and missed detection are marked with green, yellow and red circles respectively. Figure 6 As shown in , in scenes with simple background structures or high local signal-to-noise ratio, most methods can detect most targets, but detection of low-brightness targets at the edge of the image is still challenging. Figure 6 In the scenario shown, only the present invention can accurately detect all targets, and the results are highly consistent with the true labels. When the background is complex or the targets are densely distributed, other methods generally miss detection and make false detections. For example, in Figure 6 While other methods clearly missed detections, our method successfully detected all small targets. Even in the presence of highly complex noise interference, such as mountain contours, or low-contrast cloud cover, our method still achieved accurate detection without false positives. These results fully demonstrate the stability and superiority of our method in complex scenarios.

[0088] In order to further demonstrate the computational efficiency advantage of Mamba over Transformer and verify the proposed 、 as well as To investigate the effectiveness of the proposed model and its internal design, we conducted systematic ablation experiments and averaged the results. As shown in Table 3, replacing the Mamba-based global branch in the encoder with a Vision Transformer (ViT)-based branch, the two models achieved comparable performance in detection accuracy. However, the ViT-based model was significantly more computationally intensive and parameter-intensive than the Mamba-based design. As shown in Table 4, using UNet as the base model, the average IoU improvement was 4.78% after all modules were introduced. The average improvement was 2.11%. The average reduction With the gradual addition of each module, the model performance continues to improve, fully verifying the effectiveness of the proposed structural design. To understand the role of the module, we designed several sets of ablation experiments to analyze the contribution of various differential convolutions within it to infrared small target detection. As shown in Table 5, By introducing center difference (CDC), angle difference (ADC), horizontal difference (HDC) and vertical difference (VDC) convolution, it is used to extract subtle pixel differences and edge gradient information between the target and the background. By removing different types of differential convolution kernels in turn to construct different variants, the experimental results show that with the gradual introduction of differential convolution, the model has better performance in IoU and The performance of indicators such as and continues to improve, indicating that capturing target-background pixel differences and edge gradient prior information can significantly enhance the expressiveness of target features. In order to verify the effectiveness of the module, we conducted multiple sets of ablation experiments to examine the combination of depth-wise separable convolution branches of different scales with Mamba. As shown in Table 6, the experiments include single-scale variants that only retain 1×1, 3×3 or 5×5 branches, and multi-scale fusion variants that combine branches of different scales (such as 1×1&3×3, 1×1&5×5, 3×3&5×5). The results show that each scale branch can effectively model background information under its corresponding receptive field, and the combination of multi-scale branches further improves the detection performance. In particular, after integrating branches of all three scales, the model performs best in all indicators, verifying the effectiveness and complementarity of the multi-scale background information aggregation strategy. To further evaluate The effects of spatial activation strategy and channel filtering on the fusion of global and local features in the module were investigated. We constructed multiple network variants, removing one or both of the attention mechanisms, and analyzed their performance differences. As shown in Table 7, when both strategies were removed at the same time, the model performance dropped significantly, with an average IoU reduction of 2.35%. decreased by 1.12%, The average increase is 2.60, indicating that the proposed dual attention mechanism has a significant effect in enhancing the fusion of global and local features and improving the overall detection performance.

[0089] Table 3 Comparison of computational efficiency between Mamba and Transformer. The best results are in bold.

[0090]

[0091] Table 4. Average results of arbitrary ablation of the proposed module on three datasets. The best results are in bold.

[0092]

[0093] Table 5. Further ablation results of the DCCB module. The best results are in bold.

[0094]

[0095] Table 6 Further ablation results of the MCMB module, the best results are in bold

[0096]

[0097] Table 7 Further ablation results of the FFDM module, the best results are in bold

[0098]

[0099] The above content is a further detailed description of the present invention in conjunction with specific embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art of the art to which the present invention belongs, they can also make several substitutions or modifications to these described embodiments without departing from the concept of the present invention, and these substitutions or modifications should be considered to fall within the scope of protection of the present invention. Parts of the present invention that are not described in detail belong to the common technology of those skilled in the art.

Claims

1. An infrared small target detection method integrating local priors and multi-scale global background, characterized by: The steps include: Step 1: Obtain image data containing infrared images and their corresponding binary mask labels and perform preprocessing; Step 2: Construct an object detection model with an encoder-decoder architecture. The encoder includes a parallel local detail prior mining branch and a multi-scale global background perception branch, and the decoder includes a progressive feature fusion decoding branch. The preprocessed image data is subjected to step-by-step feature extraction and decoding to obtain detection results. Step 3: Introduce a weighted deep supervision mechanism in the object detection model training: set auxiliary prediction outputs in multiple decoding layers and calculate weighted losses separately to form a multi-level supervision signal.

2. The infrared small target detection method integrating local prior and multi-scale global background according to claim 1 is characterized in that: The target detection model is specifically implemented as follows: Step 2.1: Construct a local detail prior mining branch, which consists of Layer details capture convolutional blocks The branch is composed of two branches in series, which extracts the prior information of target-background pixel contrast difference and target edge gradient change. The input of this branch is infrared image. layer Then output a local feature map , ; Step 2.2: Construct a multi-scale global background perception branch, which consists of Layer multi-scale Mamba blocks The branch accepts the infrared image as input, which is then expanded into blocks and then input into the series. , No. layer After the block merging operation, a global feature map is output ; Step 2.3: Construct a progressive fusion decoding branch, which consists of Layer feature fusion decoding module The cascaded structure gradually restores the spatial resolution and generates the target mask; layer Receive and fuse a feature map from the multi-scale global background perception branch step by step , a feature map of a local detail prior mining branch and a lower layer fusion decoding output feature map , among which layer Received lower layer fusion decoding output is by and After processing through splicing, basic convolution block, SE attention mechanism and another basic convolution block, a fused decoding feature map of this layer is output. .

3. The infrared small target detection method integrating local prior and multi-scale global background according to claim 2 is characterized in that: The specific implementation process of the local detail prior mining branch is as follows: This branch is The layer detail-capturing convolution blocks are connected in series; the detail-capturing convolution block first expands the feature dimension through a basic convolution block, and then the feature input is parallel to an original convolution kernel and four differential convolution kernels. The convolution results are added and fused, and then processed by a batch normalization and ReLU activation function to output the local feature map.

4. The infrared small target detection method integrating local prior and multi-scale global background according to claim 3 is characterized in that: The four differential convolution kernels are: center differential convolution CDC, angle differential convolution ADC, horizontal differential convolution HDC and vertical differential convolution VDC; CDC and ADC extract the pixel difference prior between the target and the background, and HDC and VDC are combined to extract the target edge gradient change prior; The basic convolution block is composed of a cascade of convolutional layers, batch normalization layers, and ReLU activation function layers.

5. The infrared small target detection method integrating local prior and multi-scale global background according to claim 4 is characterized in that: The specific implementation process of the multi-scale global background perception branch is as follows: This branch is The multi-scale Mamba block is composed of a series of multi-scale Mamba blocks. A depthwise separable convolution block is first used inside the multi-scale Mamba block for channel expansion. Then a visual state space block is used to model long-range dependencies through the state space mechanism of the two-dimensional selective scanning operator. State recursion is performed on the input feature map in the horizontal and vertical directions respectively. Then, global state perception is completed after a linear layer and a GELU activation function. After completing global state perception, the visual state space block is connected in series with three parallel depthwise separable convolution blocks of different scales. The convolution kernel size increases step by step. The three parallel output feature maps are fused by addition and finally integrated using another linear layer to obtain the output.

6. The infrared small target detection method integrating local prior and multi-scale global background according to claim 5 is characterized in that: The specific implementation process of the progressive feature fusion decoding branch is as follows: This branch is The layer feature fusion decoding module is connected in series, which is used to combine the feature maps from the global branch in the decoding stage. , feature map of local branch And the fusion decoding output feature map of the lower layer Carry out level-by-level fusion; specifically: layer, Will 、 and After splicing, the basic convolution block is used to reduce the channel to obtain the splicing features , and then use the spatial activation strategy and channel filtering strategy to process them respectively. Finally, the global features of the spatial activation strategy and the local features of the channel filtering strategy are added after the basic convolution block processing dimensions, and are processed by layer normalization and ReLU activation function to obtain , and then passed to the upper layer decoding module to achieve step-by-step feature integration from coarse to fine.

7. The infrared small target detection method integrating local prior and multi-scale global background according to claim 6 is characterized in that: The spatial activation strategy and channel filtering strategy are specifically implemented as follows: the spatial activation strategy is used to stitch features. The average pooling and maximum pooling operations are applied to the channel dimension respectively. After splicing, the spatial attention map is extracted by convolution and SigMoid activation function, and multiplied with the global feature map pixel by pixel. At the same time, the channel filtering strategy is to use the spliced ​​feature to extract the spatial attention map. Perform global average pooling and maximum pooling operations, input to the shared multi-layer perceptron and add them, then process them through SigMoid to generate channel attention weights, and compare them with the local feature map Multiply channel by channel.

8. The infrared small target detection method integrating local prior and multi-scale global background according to claim 7 is characterized in that: The weighted depth supervision mechanism sets up auxiliary output channels at each level of the decoder to supervise and guide the intermediate feature maps of different semantic depths. Specifically, each decoding layer is connected to a channel composed of The prediction head, consisting of convolution and sigmoid activation functions, generates a small target probability map whose size matches the label as an auxiliary output. This auxiliary output is compared with the true label map for loss calculation, using a weighted combination of soft intersection over union (SoftIoU) and binary cross entropy (BCE) to measure the model's target segmentation capability at different scales.

Images