Infrared small target detection method integrating local priors and multi-scale global background

Abstract

This invention discloses an infrared small target detection method that integrates local priors and multi-scale global background. The method first acquires image data containing infrared images and their corresponding mask labels, and then preprocesses it. Next, it constructs a target detection model with an encoder-decoder architecture. The encoder includes parallel local detail prior mining branches and multi-scale global background perception branches, extracting features step-by-step from the preprocessed image data. The decoder includes a progressive feature fusion decoding branch, inputting features from each level of the encoder's dual branches into the decoder branch for step-by-step decoding to obtain the detection result. Finally, a weighted depth supervision mechanism is introduced during training, setting auxiliary prediction outputs at multiple decoding layers and calculating weighted losses. This invention addresses the problems of insufficient local detail modeling, insufficient multi-scale global background perception in Mamba methods, and difficulties in fusing global and local features in existing methods, thereby improving the detection accuracy of infrared small targets.

Description

Technical Field

[0001] This invention belongs to the field of infrared small target detection technology, specifically relating to an infrared small target detection method that integrates local priors and multi-scale global background. Background Technology

[0002] Infrared imaging equipment boasts advantages such as all-weather operation, strong anti-interference capabilities, and the ability to detect camouflaged targets. However, under complex backgrounds and long-distance imaging conditions, small targets in infrared images are typically extremely small in size, have weak texture, and low contrast, making them easily submerged by background noise, posing a significant challenge to automatic detection. Since small targets occupy only a few pixels, traditional bounding box regression-based methods are not robust to them and are prone to missed detections or false alarms. Therefore, current research generally models infrared small target detection as an image segmentation problem.

[0003] Existing infrared small target detection methods mainly include traditional methods, CNN-based methods, Transformer-based methods, and a few 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, and generally suffer from sensitivity to parameters and poor generalization ability, making them difficult to adapt to complex and varied real-world backgrounds. CNN-based methods automatically learn feature representations through end-to-end training, but due to limited receptive fields, they struggle to model large-scale contextual information, leading to unstable target-level detection performance and a tendency for false alarms or missed detections. Transformers introduce a global self-attention mechanism, enhancing long-distance dependency modeling capabilities, but they have high computational overhead, insufficient real-time performance, and their expressive power is easily degraded in infrared images characterized by weak signals and limited texture. Mamba, as a state-space modeling framework proposed in recent years, combines linear computational complexity with strong global modeling capabilities, showing promising prospects 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 background and gradient changes at the target edges, resulting in weak feature responses for small targets; second, Mamba's fixed-scale image segmentation method limits its ability to perceive multi-scale backgrounds, making it difficult to adapt to the distribution characteristics of targets of different scales in complex backgrounds; finally, there is a semantic discrepancy between the global semantic information and local detail features modeled by Mamba, making alignment and fusion difficult and affecting the accurate localization and boundary reconstruction of targets.

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

[0005] To address the aforementioned problems, the present invention aims to solve the issues of insufficient local prior information capture, insufficient multi-scale global background context awareness, and difficulties in fusing global and local features when applying Mamba to infrared small targets. The invention proposes a method for detecting infrared small targets that integrates local prior information and multi-scale global background. To achieve this objective, the specific technical solution adopted by the present invention is as follows:

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

[0007] Step 1: Obtain PNG format image data containing infrared images and their corresponding binary mask labels, and perform preprocessing, including random rotation, Gaussian processing, and uniform cropping of the original data.

[0008] Step 2: Design the neural network model of the infrared small target detection method. The model is an encoder-decoder architecture, in which the encoder contains a parallel local detail prior mining branch and a multi-scale global background perception branch, 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 at each level of the encoder's two branches are input into the decoder branch to obtain the detection result step by step.

[0009] Step 2.1: Construct a local detail prior mining branch, which consists of... Layer details can capture convolutional blocks ( This branch is constructed in series and used to extract prior information such as target-background pixel contrast differences and target edge gradient changes. The input to this branch is an infrared image. layer A local feature map will then be output. .

[0010] Step 2.2: Construct a multi-scale global background perception branch, which consists of... Multi-scale Mamba blocks ( This branch is constructed in series to achieve the perception and fusion of background semantic information at different scales. The input to this branch is an infrared image, which is then processed through patch expanding (PE) before being input into the concatenated array. , No. layer The output will then undergo a patch merging (PM) operation to produce a global feature map. .

[0011] Step 2.3: Construct a progressive fusion decoding branch, which consists of... Layer feature fusion decoding module ( The process involves concatenating elements to progressively restore spatial resolution and generate a target mask; the second... layer The feature maps from the multi-scale global background perception branch are received and fused step by step. Feature map of a local detail prior mining branch A fusion decoding output feature map of a lower layer , of which layer Required lower-layer fusion decoding output It is by... and The output is a fused decoded feature map of this layer, obtained by concatenation, basic convolutional blocks, a Squeeze-and-Excitation (SE) attention mechanism, and another basic convolutional block. .

[0012] Step 3: A weighted deep supervision mechanism is introduced during training. This involves setting auxiliary prediction outputs in multiple decoding layers and calculating weighted losses for each layer, forming multi-level supervision signals to accelerate gradient backpropagation. The loss calculation uses a joint loss function consisting of Soft IoU loss and Binary Cross-Entropy loss (BCE). The former is used to optimize the contour overlap and segmentation consistency of the target region, while the latter is used to improve 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 as follows: The specific scheme of the local detail prior mining branch in step 2 is as follows: This branch is mainly composed of four layers of detail-capturing convolutional blocks connected in series. Inside each detail-capturing convolutional block, a basic convolutional block first expands the feature dimension. Then, the feature input is parallel to an original convolutional kernel and four differential convolutional kernels. The four differential convolutional kernels are: Central 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 gradient change prior of the target edge. The results of each convolution are added and fused, then processed by batch normalization and a ReLU activation function to output a local feature map. The basic convolutional block is composed of convolutional layers, batch normalization layers, and ReLU activation function layers cascaded together.

[0015] A further technical solution is as follows: The multi-scale global background perception branch in step 3 is specifically designed as follows: This branch consists of four multi-scale Mamba blocks connected in series. Within each multi-scale Mamba block, a depthwise separable convolutional block is first used for channel expansion. Then, a classic Visual State Space Block (VSS Block) is used to model long-range dependencies through the state-space mechanism of the 2D Selective Scan Operator (SS2D). State recursion is performed on the input feature map along both the horizontal and vertical directions. Global state perception is then achieved after passing through a linear layer and the GELU activation function. After completing global state perception, to improve the model's adaptability to targets and background structures at different scales, three parallel depthwise separable convolutional blocks of different scales are connected in series after the visual state space block. The kernel size increases progressively. The three parallel output feature maps are fused by addition, and finally, another linear layer is used to integrate them to obtain the output. This output integrates a high-order representation of multi-scale global background information.

[0016] A further technical solution is as follows: the progressive feature fusion decoding branch scheme in step 4 is as follows: this branch is mainly composed of four layers of feature fusion decoding modules connected in series. This module is used to process the feature map from the global branch during the decoding stage. Feature diagram of local branches The feature map is fused with the lower layer and decoded. A step-by-step fusion process is performed to improve the consistency of feature representation and the accuracy of target localization. Specifically, in the first... layer, Will , and After splicing, channel reduction is performed using basic convolutional blocks to obtain the spliced ​​features. Subsequently, spatial activation and channel filtering strategies were applied to mitigate the differences in scale and semantic level between global semantic features and local detail features. Finally, the global features obtained through spatial activation and the local features obtained through channel filtering were processed through the basic convolutional block dimensions and then summed. The results were then processed using layer normalization and the ReLU activation function to obtain... Then it is passed to the next layer of decoding module to achieve progressive feature integration from coarse to fine.

[0017] A further technical solution is as follows: the specific schemes of the spatial activation strategy and the channel filtering strategy are as follows: the spatial activation strategy applies to the splicing features Average pooling and max pooling operations are applied along the channel dimension, and the concatenated arrays are then processed by convolution and a SigMoid activation function to extract the spatial attention map, which is then combined with the global feature map. Pixel-by-pixel multiplication. Simultaneously, the channel filtering strategy modifies the stitching features. Global average pooling and max pooling operations are performed, the inputs are fed into a shared multilayer perceptron and summed, then processed by SigMoid to generate channel attention weights, which are then compared with local feature maps sequentially. Channel multiplication.

[0018] A further technical solution is as follows: In step 5, the specific scheme of the weighted depth supervision mechanism is: by setting an auxiliary output channel at each level of the decoder, it is used to supervise and guide the intermediate feature maps of different semantic depths. More specifically, each decoding layer connects to a channel after its output... A prediction head consisting of convolutional and sigmoid activation functions generates a small target probability map with a size matching the label as an auxiliary output. This auxiliary output is compared with the ground truth label map to calculate the loss, using a weighted combination of soft IoU loss and binary cross-entropy loss (BCE) to measure the model's target segmentation ability at different scales.

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

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

[0021] The multi-scale global background context-aware branch of this invention has the ability to model long-range dependencies in the image space dimension. By integrating contextual semantic information through a multi-scale depth-separable convolutional structure, it 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 this invention employs 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 this invention, while maintaining detection accuracy, alleviates the computational complexity problem caused by using Transformer by introducing Mamba for global modeling, thereby improving the feasibility of the method in engineering applications. Attached Figure Description

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

[0025] Figure 2 This refers to the prior information of local details and target information in a multi-scale background as described in this invention;

[0026] Figure 3 This is a diagram of a convolutional block structure with captureable details, as proposed in this invention.

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

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

[0029] Figure 6 This is a visual comparison of the segmentation results of the present invention and representative methods. Detailed Implementation

[0030] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but do not limit the present invention in any way. It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

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

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

[0033] In this embodiment, three publicly available infrared small target datasets were selected: (1) NUAA-SIRST dataset: This dataset is the first publicly available infrared small target dataset, containing 427 images and a total of 480 target instances. The images in 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, wilderness, 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. This invention adopts the default partitioning ratio of the dataset proposer, where the training set test set partitioning ratio of NUAA-SIRST and NUDT-SIRST is 1:1, while the training set test set partitioning ratio of IRSTD-1k is 4:1. The data were uniformly subjected to common data augmentation operations such as random rotation, Gaussian processing, and cropping, and finally became .

[0034] Step 2: Design the neural network model of the infrared small target detection method. The model is an encoder-decoder architecture, in which the encoder contains a parallel local detail prior mining branch and a multi-scale global background perception branch, 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 at each level of the encoder's two branches are input into the decoder branch to obtain the detection result step by step.

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

[0036] In this embodiment, the branch consists of four layers of detail-capturing convolutional blocks ( This branch is constructed by concatenation and is used to extract prior information such as target-background pixel contrast differences and target edge gradient changes, thereby enriching the target feature representation and enhancing the discernibility of small infrared targets against weak texture and low contrast backgrounds. The input to this branch is an infrared image. layer A local feature map will then be output. To further illustrate the aforementioned prior information, such as... Figure 2 As shown, the three images in the first row from left to right represent the original image, the prior information of pixel differences between the target and the background, and the prior information of the target edge gradient, respectively. In the original image, the target is extremely small and its feature information is weak. By introducing the above prior information, the feature representation of the small target can be enriched.

[0037] like Figure 3 As shown, Internally, the feature dimensions are first expanded using a basic convolutional block. Then, the feature input is processed in parallel by a primary convolutional kernel and four differential convolutional kernels: Central Differential Convolution (CDC), Angle Differential Convolution (ADC), Horizontal Differential Convolution (HDC), and Vertical Differential Convolution (VDC). CDC and ADC are used to extract prior pixel differences between the target and background. HDC and VDC are combined to extract prior gradient changes at the target edges. The results of each convolution are summed and fused, then processed by batch normalization and a ReLU activation function to output a local feature map. The basic convolutional block consists of cascaded convolutional layers, batch normalization layers, and ReLU activation function layers. The four differential convolutions are derived from the primary convolutional kernel, and the primary convolutional kernel and the four differential convolution kernels are defined as follows:

[0038]

[0039]

[0040]

[0041]

[0042]

[0043] in Represents the original convolution kernel. , , and These represent the horizontal difference convolution kernel, the vertical difference convolution kernel, the central difference convolution kernel, and the angular difference convolution kernel, respectively. Let the weights be denoted as . Specifically, let the input features and output features be denoted as . and ,in Then a complete The feature processing procedure of a module can be defined as follows:

[0044]

[0045] in Represents the activation function. This represents combining the original convolution with four different convolution types in parallel, and then concatenating them with a batch-normalized difference-capture convolution. This represents the basic convolutional block.

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

[0047] In this embodiment, the branch consists of four layers of multi-scale Mamba blocks ( This branch, constructed in concatenation, is used to perceive and fuse semantic information from different scales of background. It effectively enhances the modeling ability for complex background interference and the ability to capture multi-scale background context dependencies, while maintaining linear computational complexity. The input to this branch is an infrared image, which is then processed through patch expanding (PE) before being input into the concatenated... , No. layer The output will then undergo a patch merging (PM) operation to produce a global feature map. To further illustrate the aforementioned multi-scale global background information, such as... Figure 2As shown, the four images in the second row from left to right represent the target rendering effect against a small-scale background, a medium-scale background, a large-scale background, and a multi-scale background, respectively. It can be seen that as the background scale increases, the target localization 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 is possible to retain the detailed information of the target and suppress the interference caused by the large background at a small scale, while obtaining more sufficient contextual semantics at a larger scale. This alleviates the noise introduced by the single-scale background and achieves accurate and robust target localization.

[0048] like Figure 4 As shown, Internally, a depthwise separable convolutional 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 its two-dimensional selective scan operator (SS2D). State recursion is performed on the input feature map along both the horizontal and vertical directions, thereby achieving global state awareness in the spatial dimension. Specifically, let the input feature map be: The bidirectional state modeling process in SS2D is as follows:

[0049]

[0050]

[0051]

[0052] in, and These represent the states in the row (width) and column (height) directions, respectively. The coordinates of the image after it has been divided into blocks; , are the learnable state transition matrix and input projection matrix. To output the mapping weight matrix, the fusion weights of the image in both directions can be further learned. Initial state. =0. Then, after a linear layer and the GELU activation function, global state awareness is achieved. To improve the model's adaptability to targets and background structures at different scales, this module is further connected in series with three parallel, depthwise separable convolutional blocks of different scales, with kernel sizes of: and The parallel outputs are fused through addition, and finally integrated using another linear layer to obtain the final output. Let the input features and output features be respectively... ,in Then a complete The module feature processing process can be described as follows:

[0053]

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

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

[0056] In this embodiment, the branch consists of a four-layer feature fusion decoding module ( The process involves a series of steps to gradually restore spatial resolution and generate a target mask. This branch effectively improves the consistency of feature representation and the accuracy of target localization. Layer feature fusion decoding module It will receive and fuse feature maps from a multi-scale global background perception branch at each level. Feature map of a local detail prior mining branch A fusion decoding output feature map of a lower layer (of which the fourth layer) Required lower-layer fusion decoding output It is by... and The output, obtained after concatenation, basic convolutional blocks, a Squeeze-and-Excitation (SE) attention mechanism, and another basic convolutional block, is a fused decoded feature map of this layer. .

[0057] like Figure 5 As shown, Used for feature maps from the global branch during the decoding phase. Feature diagram of local branches The feature map is fused with the lower layer and decoded. A step-by-step fusion process is performed to improve the consistency of feature representation and the accuracy of target localization. Specifically, in the first... layer, Will , and The splicing features are obtained by splicing. Subsequently, after channel reduction using basic convolutional blocks, spatial activation and channel filtering strategies are applied to mitigate the differences in scale and semantic level between global semantic features and local detail features. Finally, the global features processed by spatial activation and the local features processed by channel filtering are processed by basic convolutional blocks and then summed. Finally, layer normalization and ReLU activation are applied before passing the sum to the next layer's decoding module, achieving a progressive feature integration from coarse to fine. Let the input features be... and The output features are ,in Then a complete The module feature processing process can be described as follows:

[0058]

[0059]

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

[0061] like Figure 5 As shown, the spatial activation strategy affects the splicing features. Average pooling and max pooling operations are applied separately along the channel dimension, and the result is concatenated. Convolutional processing extracts a spatial attention map, which is then multiplied pixel-by-pixel with the global feature map, significantly enhancing the spatial response intensity of the target region and guiding the model to focus more accurately on small target regions. Simultaneously, the channel filtering strategy, through the processing of concatenated features… Global average pooling and max pooling operations are performed, and the input is fed into a shared multilayer perceptron to generate channel attention weights. These weights are then multiplied channel-by-channel with local feature maps to adaptively adjust the response intensity of each channel, enhancing semantic channels relevant to small infrared targets and suppressing redundant background interference. Let the input be the stitched features. Spatial activation strategy and channel filtering strategy The detailed processing steps can be represented as follows:

[0062]

[0063]

[0064] in Represents the activation function. Represents a multilayer perceptron. and These represent average pooling and max pooling, respectively. represent convolution, Represents splicing operation

[0065] Step 6: A weighted deep supervision mechanism is introduced during training. This involves setting auxiliary prediction outputs in multiple decoding layers and calculating weighted losses for each, forming multi-level supervision signals. This accelerates gradient backpropagation, improves convergence efficiency, and enhances the intermediate layers' ability to perceive and robust against small targets. The loss calculation uses a joint loss function consisting of Soft IoU loss and Binary Cross-Entropy loss (BCE). The former optimizes the contour overlap and segmentation consistency of the target region, while the latter improves the pixel-level classification accuracy of the target and background.

[0066] like Figure 1 As shown, specifically, each decoding layer connects to a converter after its output. A prediction head composed of convolutional layers and a sigmoid activation function generates a small target probability map with a size matching the label as an auxiliary output. This auxiliary output is compared with the ground truth label map for loss calculation, using a weighted combination of soft IoU loss and binary cross-entropy loss (BCE) to measure the model's target segmentation ability at different scales. This loss calculation process can be described as follows:

[0067]

[0068] in Represents the loss value. Represents the real mask. , , representing the mask for each output channel of the decoder, where An additional layer is added after concatenating the first five auxiliary prediction masks. The final output mask is obtained from the prediction head composed of convolution and sigmoid activation functions. During training, all masks participate in the loss calculation; during testing, only one mask is used. As a prediction result, this supervision mechanism not only accelerates network convergence and improves training stability, but also enhances the expressive power and semantic consistency of multi-layer feature maps, significantly improving the model's robustness and localization accuracy for detecting small infrared targets.

[0069] In this embodiment, during training, we built the model using PyTorch 2.0 on an NVIDIA RTX 4090 GPU equipped with 24GB of VRAM. For 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 reduced to 0.005 using a cosine annealing strategy. To achieve a smoother and more efficient optimization process, the Adagrad optimization algorithm was used for parameter updates, which is suitable for cases with sparse gradients. The entire model was trained in 400 batches with a batch size of 8. To ensure fairness in the learning process and objectivity in the comparison results, all models were trained from scratch without using pre-trained weights.

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

[0071] To verify the performance of the present invention, in this embodiment, it was evaluated and compared with several advanced methods in the art, specifically including:

[0072] (1) Traditional methods: including Top-Hat (morphological top-hat transformation 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 Association 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 Detector Network), DMFNet (Dual Encoder Multi-Stage Feature Fusion Network), PBT (Progressive Background Awareness Transformer), MiM (Nested Mamba Network), ID-UNet (Dense Connection U-Net Network), and MMLNet (Multi-Branch Inter-Learning Network).

[0074] To comprehensively evaluate the superior effects of this invention, the evaluation metrics used include:

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

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

[0077] Normalized intersection-union ratio (nIoU): nIoU is the normalized version of IoU;

[0078] F1 Score: This metric evaluates the false negative and false positive rates at the pixel level.

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

[0080] Detection rate ( ): The proportion of correctly detected targets is defined as the number of correctly predicted targets. Total target quantity The ratio. If the center point deviation of the predicted target is less than 3 pixels, the target is considered to have been correctly detected.

[0081] False alarm rate ( ): Used to measure the number of pixels that are incorrectly predicted as targets. The total number of pixels predicted as the target out of all pixels in the image The ratio;

[0082] The comparison results are shown in Tables 1 and 2. In all metrics for pixel-level segmentation and target-level detection, the deep learning-based method significantly outperforms traditional methods. This indicates that under complex conditions such as cluttered backgrounds or weak targets, traditional methods have limited effectiveness, while deep models can achieve higher accuracy through data-driven learning. Among these methods, this invention achieves the best overall performance across all metrics and datasets. On the NUAA-SIRST dataset, thanks to its local prior information mining strategy, this invention achieves the highest IoU. This strategy enhances the representation ability of small targets by modeling the difference in pixel contrast between the target and background and the saliency of target edges. On the NUDT-SIRST dataset, although DMFNet improves IoU by retaining more shallow features, its lack of global context modeling results in lower IoU. This represents a 4.07% reduction compared to the previous invention, highlighting the advantages of the multi-scale global background modeling in this invention. Even on the most challenging IRSTD-1K dataset, this invention maintains a good balance between segmentation and detection, thanks to the feature fusion strategy proposed in this invention, which effectively integrates features across different levels, narrowing the gap between global and local features, thereby improving decoding accuracy.

[0083] Table 1 shows the quantitative comparison results, comparing the present invention with existing methods on the NUAA-SIRST dataset. Best results are shown in bold.

[0084]

[0085] Table 2 shows the quantitative comparison results, comparing the present invention with existing methods on the NUDT-SISRT and IRSTD-1k datasets. Best results are shown in bold.

[0086]

[0087] To further verify the effectiveness of the present invention, such as Figure 6 As shown, several typical scenes were randomly selected, and the visual detection results were compared with those of two state-of-the-art methods. The target areas are marked with blue boxes and magnified when necessary; correct detections, false alarms, and missed detections are marked with green, yellow, and red circles, respectively. Figure 6 As shown, in scenes with simple background structures or high local signal-to-noise ratios, most methods can detect most targets, but detecting low-brightness targets located at the image edges remains challenging. It is worth noting that in... Figure 6 In the scenario shown, only this invention achieves accurate detection of all targets, and the results are highly consistent with the actual labels. When the background is complex or the targets are densely distributed, other methods generally experience missed detections and false detections. For example, in… Figure 6 In contrast to other methods, which often miss detections, this invention successfully detected all small targets. Even in situations with highly complex noise interference such as mountain outlines or low-contrast cloud cover, this invention still achieves accurate detection without false alarms. These results fully demonstrate the stability and superiority of this invention in complex scenarios.

[0088] To further demonstrate the computational efficiency advantage of Mamba over Transformer and to verify the proposed... , as well as To assess the effectiveness of the modules and their internal design, systematic ablation experiments were conducted, and the results were averaged. As shown in Table 3, replacing the Mamba-based global branch in the encoder of the model of this invention with a Vision Transformer (ViT)-based branch, both models exhibited comparable performance in detection accuracy. However, compared to the Mamba-based design, the ViT-based model had significantly higher computational and parameter requirements. As shown in Table 4, using UNet as the base model, the average IoU improved by 4.78% after introducing all modules. On average, it increased by 2.11%, while The average decreased As each module was gradually added, the model's performance continued to improve, fully validating the effectiveness of the proposed structural design. Further evaluation... To illustrate the module's function, multiple ablation experiments were designed to analyze the contributions of various differential convolutions within it to infrared small target detection. As shown in Table 5, By introducing central difference (CDC), angular difference (ADC), horizontal difference (HDC), and vertical difference (VDC) convolutions, subtle pixel differences and edge gradient information between the target and the background are extracted. Different variants are constructed by sequentially removing the various difference convolution kernels. Experimental results show that with the gradual introduction of difference convolutions, the model improves IoU and... The continuously improving performance across metrics indicates that capturing prior information about target-background pixel differences and edge gradients significantly enhances the expressive power of target features. Further evaluation is needed... To assess the module's effectiveness, multiple ablation experiments were conducted, examining the combination effect of depthwise separable convolutional branches at different scales with Mamba. As shown in Table 6, the experiments included single-scale variants retaining only 1×1, 3×3, or 5×5 branches, as well as multi-scale fusion variants combining different scale branches (e.g., 1×1 & 3×3, 1×1 & 5×5, 3×3 & 5×5). The results show that each scale branch can effectively model background information within its corresponding receptive field, and the combination of multi-scale branches further improves detection performance. In particular, after integrating all three scale branches, the model performs optimally across all metrics, validating the effectiveness and complementarity of the multi-scale background information aggregation strategy. Further evaluation... The impact of spatial activation strategies and channel filtering on global and local feature fusion was investigated. Multiple network variants were constructed, and one or both attention mechanisms were removed for each variant, with performance differences analyzed. As shown in Table 7, removing both strategies simultaneously significantly reduced model performance, with an average IoU decrease of 2.35%. Decreased by 1.12%, The average increase of 2.60 indicates that the proposed dual attention mechanism has a significant effect on enhancing the fusion of global and local features and improving overall detection performance.

[0089] Table 3 Comparison of computational efficiency between Mamba and Transformer, with the best results shown in bold.

[0090]

[0091] Table 4 shows the average arbitrary ablation results for the proposed module on three datasets, with the best results indicated in bold.

[0092]

[0093] Table 5 shows the further ablation results of the DCCB module; the best results are indicated in bold.

[0094]

[0095] Table 6 shows the further ablation results for the MCMB module; the best results are indicated in bold.

[0096]

[0097] Table 7 shows further ablation results for the FFDM module; the best results are indicated in bold.

[0098]

[0099] The above description, in conjunction with specific embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various substitutions or modifications can be made to these described embodiments without departing from the concept of the present invention, and all such substitutions or modifications should be considered within the scope of protection of the present invention. The parts of the present invention not described in detail are well-known to those skilled in the art.

Claims

1. An infrared small target detection method that integrates local priors and multi-scale global background, characterized in that, The steps include: Step 1: Obtain image data containing infrared images and their corresponding binary mask labels, and perform preprocessing; Step 2: Construct an encoder-decoder architecture target detection model. The encoder includes a parallel local detail prior mining branch and a multi-scale global background perception branch. The decoder includes a progressive feature fusion decoding branch. The detection results are obtained by progressively extracting and decoding features from the preprocessed image data. The target detection model is implemented as follows: Step 2.1: Construct a local detail prior mining branch, which consists of... Layer details can capture convolutional blocks This is a concatenated structure that extracts prior information about the contrast difference between target and background pixels and the gradient changes at the target edges; the input to this branch is an infrared image, the first... layer Then output a local feature map , ; The specific implementation of the local detail prior mining branch is as follows: This branch is composed of... The layer details can be captured by concatenating convolutional blocks; inside the detail-capturing convolutional block, the feature dimension is first expanded by a basic convolutional block, and then the feature input is parallel to a raw convolutional kernel and four differential convolutional kernels. The convolutional results are added and fused, and then processed by a batch normalization and ReLU activation function to output a local feature map. Step 2.2: Construct a multi-scale global background perception branch, which consists of... Multi-scale Mamba blocks This concatenated structure enables the perception and fusion of semantic information from different scales of background. The input to this branch is an infrared image, which is then processed through a block-expanding operation before being input into the concatenated structure. , No. layer After block merging, a global feature map is output. ; The multi-scale global background perception branch is specifically implemented as follows: This branch consists of... The system is composed of multiple Mamba blocks of different scales. Each Mamba block first uses a depthwise separable convolutional block for channel expansion, followed by a visual state space block. This block models long-range dependencies through the state space mechanism of a two-dimensional selective scanning operator, performing state recursion on the input feature map along both the horizontal and vertical directions. Then, it passes through a linear layer and a GELU activation function to complete global state perception. After completing global state perception, three parallel depthwise separable convolutional blocks of different scales are concatenated after the visual state space block. The kernel size increases progressively. The three parallel output feature maps are fused by addition, and finally, another linear layer is used to integrate them to obtain the output. Step 2.3: Construct a progressive fusion decoding branch, which consists of... Layer feature fusion decoding module The process involves cascading steps to gradually restore spatial resolution and generate a target mask; the first step... layer The feature maps from the multi-scale global background perception branch are received and fused step by step. Feature map of a local detail prior mining branch A fusion decoding output feature map of a lower layer , of which layer Received lower-layer fusion decoding output It is by... and The output is a fused decoded feature map of this layer, obtained by concatenation, basic convolutional blocks, SE attention mechanism, and another basic convolutional block. ; Step 3: Introduce a weighted deep supervision mechanism in the training of the object detection model: Set auxiliary prediction outputs in multiple decoding layers and calculate weighted losses respectively to form multi-level supervision signals.

2. The infrared small target detection method that integrates local prior and multi-scale global background according to claim 1, characterized in that, The four differential convolution kernels are: central differential convolution CDC, angular 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 convolutional block consists of a cascaded convolutional layer, a batch normalization layer, and a ReLU activation function layer.

3. The infrared small target detection method according to claim 2, which integrates local prior and multi-scale global background, 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 composed of multiple layers connected in series. This module is used to process the feature maps from the global branch during the decoding stage. Feature diagram of local branches The feature map is fused with the lower layer and decoded. To carry out step-by-step integration; specifically: in the first stage layer, Will , and After splicing, channel reduction is performed using basic convolutional blocks to obtain the spliced ​​features. Subsequently, spatial activation and channel filtering strategies were applied respectively. Finally, the global features obtained from the spatial activation strategy and the local features obtained from the channel filtering strategy were processed through the basic convolutional block dimensions and then summed. The results were then processed using layer normalization and the ReLU activation function to obtain the final product. Then it is passed to the next layer of decoding module to achieve progressive feature integration from coarse to fine.

4. The infrared small target detection method that integrates local prior and multi-scale global background according to claim 3, characterized in that, The spatial activation strategy and channel filtering strategy are specifically implemented as follows: the spatial activation strategy applies splicing features... Average pooling and max pooling operations are applied separately along the channel dimension. After concatenation, the concatenated features are processed by convolution and the SigMoid activation function to extract the spatial attention map, which is then multiplied pixel-by-pixel with the global feature map. Simultaneously, a channel filtering strategy is applied to the concatenated features... Global average pooling and max pooling operations are performed, the inputs are fed into a shared multilayer perceptron and summed, then processed by SigMoid to generate channel attention weights, which are then combined with local feature maps. Multiply each channel.

5. The infrared small target detection method according to claim 4, which integrates local prior and multi-scale global background, is characterized in that... The weighted depth supervision mechanism sets up an auxiliary output channel at each level of the decoder to supervise and guide intermediate feature maps of different semantic depths. Specifically, each decoding layer connects an auxiliary output channel after its output. The prediction head, composed of convolution and sigmoid activation functions, generates a small target probability map with a size matching the label as an auxiliary output. This auxiliary output is compared with the ground truth label map for loss calculation. A weighted combination of soft IoU loss and binary cross-entropy loss (BCE) is used to measure the model's target segmentation ability at different scales.

Images