Infrared weak and small target segmentation method based on depth expansion and adaptive weighting

Abstract

The invention discloses an infrared dim and small target segmentation method based on depth expansion and adaptive weighting, and belongs to the technical field of infrared dim and small target segmentation. The method comprises the following steps: modeling a segmentation task as a sparse-low rank matrix decomposition problem, constructing a multi-stage trainable module containing LBRM, STEM and WEM by using a depth expansion network, a dual-path complementary attention mechanism is introduced to fuse local and global features, a self-adaptive weighting matrix is constructed in combination with a local structure difference weight and a sparse enhancement weight, SoftIoU, MSE and Sparse combination loss function optimization is adopted, the method realizes high-precision segmentation on NUDT-SIRST and SIRST-Aug data sets, the detection accuracy and model interpretability under a complex background are improved, and the method is suitable for large-scale popularization and application. The method is suitable for night monitoring, maritime rescue and other scenes.

Description

Technical Field

[0001] This invention relates to the field of infrared weak target segmentation technology, specifically to an infrared weak target segmentation method based on depth unfolding and adaptive weighting. Background Technology

[0002] Infrared weak target segmentation is an important application of infrared imaging technology, widely used in scenarios such as nighttime monitoring, traffic security, disaster relief, and forest fire prevention. It can also enable accurate detection of personnel and vessels in maritime rescue and identification of pedestrians and animals in autonomous driving.

[0003] Existing infrared weak target segmentation methods are mainly divided into two categories: traditional methods based on manually generated features and data-driven deep learning methods. However, both have limitations in complex scenes. Traditional methods rely on manually designed feature operators such as edge detection, local contrast, and texture extraction. They can achieve target localization in simple backgrounds, but they are prone to false positives and false negatives in complex scenes with strong noise and low target-background contrast. Furthermore, they lack adaptive mechanisms, require manual adjustment of hyperparameters, have poor robustness, and are difficult to cope with interference from sea waves, dynamic clouds, and other factors.

[0004] In recent years, significant progress has been made in infrared weak target segmentation methods based on deep learning. Convolutional neural networks (CNNs) have improved the accuracy of target-background separation in complex scenes through their powerful feature extraction capabilities; advanced architectures such as Transformer have demonstrated potential in target segmentation by achieving global modeling through self-attention mechanisms. However, CNNs struggle to capture long-distance pixel dependencies, limiting global relationship modeling; Transformer's global computation mechanism wastes computational resources because weak targets occupy only a small number of pixels, and its "black box" design lacks clear domain knowledge, making the decision-making process difficult to interpret and high-confidence errors difficult to trace. Furthermore, it is highly dependent on large-scale datasets, limiting its application.

[0005] The main shortcomings of current infrared weak target segmentation technology include the following aspects: First, traditional methods have limited detection accuracy and stability, making it difficult to cope with noise and interference in complex backgrounds; second, although deep learning models have advantages in feature learning, they suffer from poor interpretability, high computational cost, and strong dependence on datasets; finally, existing deep unfolding networks that combine optimization theory and deep learning still need improvement in adaptive weight adjustment and multi-dimensional feature fusion, making it difficult to balance segmentation accuracy, interpretability, and generalization ability.

[0006] Therefore, there is an urgent need to develop an infrared weak target segmentation model that combines high accuracy, strong interpretability, and high efficiency to achieve accurate segmentation in complex scenes, while reducing computational complexity and improving robustness and generalization ability in practical applications. This will not only promote the in-depth application of infrared imaging technology in multiple fields, but also provide important technical support for target detection tasks in complex environments. Summary of the Invention

[0007] The purpose of this invention is to solve the problems of existing technologies and provide an infrared weak target segmentation method based on depth unfolding and adaptive weighting that can solve the problems of insufficient interpretability, poor adaptability to complex scenes, rigid weight adjustment mechanism, and difficulty in balancing computational efficiency and performance in existing infrared weak target segmentation models.

[0008] To achieve the above objectives, the present invention is implemented through the following technical solution:

[0009] This invention discloses an infrared weak target segmentation method based on depth unfolding and adaptive weighting, which includes the following steps:

[0010] S1, take the input infrared image as the observation item and establish a sparse-low-rank decomposition model to express the separation of background and target components;

[0011] S2, the iterative solution process of the sparse-low-rank decomposition is transformed into a trainable multi-stage network through deep expansion. Each stage includes a low-rank background reconstruction module (LBRM), a sparse target estimation module (STEM), and a weight estimation module (WEM); wherein, the weight estimation module (WEM) generates sparse enhancement weights W. SE Combined with the local structural difference weights W calculated from the structural tensor SD The adaptive weighting matrix W = W is obtained. SD ⊙W SE This matrix is ​​used to constrain the sparse estimation of the Sparse Target Estimation Module (STEM) and incorporates a dual-path complementary attention mechanism in the Low-Rank Background Reconstruction Module (LBRM) to enhance target-background separation.

[0012] S3, the initial depth unfolding network is trained end-to-end based on the infrared weak target dataset. During the training process, a combined loss function consisting of SoftIoU loss, MSE loss and Sparsity loss is used to optimize the network parameters, so as to achieve simultaneous optimization of target accuracy, structure fidelity and sparsity characteristics.

[0013] Among them: SoftIoU loss is used to improve the segmentation overlap accuracy of small target regions, MSE loss is used to maintain the consistency between input and output images to avoid loss of details, and Sparsity loss is used to enhance target sparsity and suppress background noise interference. Through the joint constraint of the three, the target accuracy, structural fidelity and sparsity characteristics are optimized simultaneously, thereby ensuring the complete preservation and high-precision segmentation of weak targets in complex backgrounds.

[0014] S4. Use the trained infrared weak target segmentation model to perform inference segmentation on the infrared image to obtain a segmentation result image containing the separated weak targets.

[0015] Preferably, each stage of the deep unfolded network includes a low-rank background reconstruction module (LBRM), a sparse target estimation module (STEM), and a weight estimation module (WEM), wherein each module is used for low-rank background modeling, sparse target estimation, and adaptive weight update, respectively, to achieve the separation of background component B and sparse target component T in the observed image F = T + B.

[0016] Preferably, the low-rank background reconstruction module (LBRM) replaces the traditional optimization process with an end-to-end learnable approach, including steps such as convolutional layer mapping, introduction of dual-path parallel attention module (DCA), batch normalization and ReLU transformation, feature extraction of convolutional stacking units, and convolutional channel compression to output the background estimation image.

[0017] Specifically, the steps include: first, using convolutional layers to map the input to the feature space; then introducing a dual-path parallel attention module (DCA); followed by batch normalization and ReLU to achieve nonlinear transformation; then extracting deep features through stacked units of convolution-normalization-activation; and finally outputting the background estimation image by compressing the number of channels through convolution.

[0018] Preferably, the dual-path parallel attention module (DCA) includes: a first path combining channel attention and spatial attention to highlight the target region; specifically, channel attention generates a weight vector through one-dimensional convolution after global average pooling and max pooling, and spatial attention constructs a two-dimensional weight map based on channel responses; the second path performs global context modeling based on a query-key-value mechanism, calculates the importance of space and channels, and integrates long-distance dependency information; the two paths are calculated in parallel and then fused for output to enhance the accuracy and robustness of background modeling.

[0019] Preferably, the Low-Rank Background Reconstruction Module (LBRM) is a learnable subnetwork designed for efficiently extracting background information from the optimization model. To improve background representation capabilities while maintaining computational efficiency, the LBRM expands the traditional background estimation process into a learnable module, forming an end-to-end background reconstruction mechanism. Structurally, it first extracts initial features through 3×3 convolutions and ReLU, then fuses local spatial background information through multiple 3×3 convolutional blocks, and finally outputs reconstructed background features by adjusting the number of channels through convolution.

[0020] Specifically, the Low-Rank Background Reconstruction Module (LBRM) first applies a 3×3 convolution followed by ReLU activation to the input features for initial feature extraction. Then, it captures and fuses local spatial background information through multiple consecutive 3×3 convolutional blocks. Finally, a convolutional layer adjusts the number of channels to match the input channels, outputting the reconstructed background features. The convolutional operation is used to extract local texture and structural information from the background, and the ReLU activation produces a non-linear suppression effect, effectively suppressing target interference and noise. Compared to traditional explicit background solving methods, the Low-Rank Background Reconstruction Module (LBRM) significantly improves computational efficiency and the adaptability of background reconstruction while maintaining the structural constraints and interpretability of background modeling.

[0021] Preferably, the weight estimation module (WEM) includes an input convolutional layer, several convolution-normalization-activation units, and an output mapping layer. The input convolutional layer is used to extract initial features, the convolution-normalization-activation units are used for multi-scale texture feature extraction and nonlinear mapping, and the output mapping layer transforms the features into sparse enhancement weights W with the same size as the input image. SE The module can adaptively adjust the weights according to the sparse target distribution at different stages, so as to dynamically optimize the salience of the target region during the iterative update process.

[0022] Preferably, in the combined loss function, SoftIoU loss is used to optimize segmentation overlap accuracy, MSE loss preserves image details, and Sparsity loss enhances target sparsity;

[0023] The total loss function is:

[0024] L totalLoss =α·SoftIoU_Loss+β·MSE_Loss+γ·SPARSITY_Loss;

[0025] Here, α, β, and γ are hyperparameters used to adjust the relative weights of each loss term during training.

[0026] Preferably, the infrared weak target dataset is NUDT-SIRST and SIRST-Aug, the former containing various complex backgrounds and low target signal-to-noise ratio, and the latter being an enhanced version to improve the model's generalization ability.

[0027] Preferably, the W SD The calculation formula is: Where h is the weight sensitivity hyperparameter; d(x,y)=|L1(x,y)-L2(x,y)| represents the difference in structural strength between the main direction and the sub-direction at pixel (x,y), and d(x,y) represents the difference in structural strength between pixels.

[0028] Preferably, the update formula for the Lagrange multiplier Y is also included:

[0029] Y k+1 =Y k +(FB k+1 -T k+1 ) / μ k To balance data consistency constraints; where Y k It is the Lagrange multiplier at the k-th iteration, used to balance data consistency constraints; Y k+1 It is the updated Lagrange multiplier at the (k+1)th iteration; F is the original infrared image; B k+1 This is the low-rank background estimation result at the (k+1)th iteration; T k+1 This is the sparse target estimation result at the (k+1)th iteration; μ k It is the step size coefficient at the k-th iteration, used to control the update magnitude of the Lagrange multipliers and ensure the stability and convergence of the optimization process.

[0030] Beneficial effects: This invention reduces the number of model parameters (only 0.681M) and improves computational efficiency through lightweight feature extraction and modular design; the DCA mechanism enhances the accuracy and robustness of target segmentation in complex backgrounds; adaptive weights and composite loss functions improve target boundary accuracy and the ability to identify weak targets. Experiments show that the model achieves mIoU of 94.37% and 74.13% on the NUDT-SIRST and SIRST-Aug datasets, respectively, and F1-scores of 97.10% and 85.15%, respectively, making it suitable for resource-constrained real-world scenarios. Attached Figure Description

[0031] Figure 1 This is a flowchart of the present invention.

[0032] Figure 2 This is a network model structure diagram of the present invention.

[0033] Figure 3 This is a diagram of the DCA network structure of the present invention.

[0034] Figure 4 This is a schematic diagram of the segmentation result of the present invention. Detailed Implementation

[0035] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0036] Current technical challenges: Infrared segmentation of weak targets is widely used in scenarios such as nighttime surveillance and disaster relief, but existing methods have limitations: traditional methods rely on manual feature extraction, resulting in poor robustness in complex scenarios; while deep learning methods improve feature extraction capabilities, convolutional networks lack global modeling capabilities, and Transformers suffer from computational waste and poor interpretability. Current technologies also face challenges such as insufficient interpretability, poor adaptability to complex scenarios, rigid weight adjustment, and difficulty in balancing computational efficiency and performance.

[0037] To address the problems in existing technologies, the technical solution / principle of this invention is as follows: This invention models infrared weak target segmentation as a sparse-low-rank matrix factorization problem, employing a deep unfolding module stacking structure. The core modules include a low-rank background reconstruction module (LBRM), a sparse target estimation module (STEM), and a weight estimation module (WEM). A DCA attention mechanism is introduced, combined with a loss function integrating SoftIoU, MSE, and Sparsity. Specifically, this includes: training and validation based on the NUDT-SIRST and SIRST-Aug datasets; decomposing the image into a low-rank background B and sparse targets T, and iteratively optimizing through a multi-stage deep unfolding network; introducing a DCA mechanism into the LBRM to fuse local and global features; and using a composite loss function to optimize the model.

[0038] Explanation of each module:

[0039] LBRM module: First, it maps the input to the feature space through convolutional layers, and then introduces... Figure 3The DCA module shown enhances feature representation. The DCA module comprises two parallel paths: the first path collaboratively integrates channel and spatial attention mechanisms, guiding the model to focus on detailed features such as target edges and textures by strengthening the modeling of key channels and spatial focusing on the target region, thereby effectively enhancing the local representation capability of small targets. The channel attention module integrates global average pooling and max pooling features, using one-dimensional convolution to capture the dependencies between channels, thereby identifying and enhancing key channels important to the target. The spatial attention module groups feature maps based on channel response intensity, extracting spatial information from important and unimportant channels separately and assigning different weights to highlight significant spatial feature responses near the target region and edges. This path significantly improves sensitivity to details under complex background conditions such as low contrast and blurred edges, enhancing the response intensity and texture recovery capability of the target region. The second path focuses on modeling the global contextual information of the image. This module first transforms the input features into three sets of information through convolution: Query represents "which channels are worth paying attention to," Key represents "which locations are worth paying attention to," and Value is the actual information contained in the image. The module performs a softmax operation on the Key in the spatial dimension to measure the importance of different spatial locations, and simultaneously performs a softmax operation on the Query in the channel dimension to reflect the importance of different channels. Then, the module multiplies the Key and Value to obtain concise global context information. This global information is then interacted with the Query to map back to various spatial locations, enhancing the global context information for long-distance relationships. These two paths respectively operate on key feature selection in the channel dimension, important information extraction in the spatial dimension, and context integration in the global dimension, achieving complementary fusion across multiple scales and perspectives, significantly improving the localization accuracy in infrared small target segmentation tasks.

[0040] The STEM module is designed to optimize the approximate solution of the sparse target subproblem in the model. It unfolds the traditional iterative process into a lightweight, learnable subnetwork, achieving end-to-end sparse target estimation. Structurally, it first extracts initial features through 3×3 convolutions and ReLU, then concatenates multiple 3×3 convolutional blocks to capture local spatial information, and finally outputs a single-channel sparse target estimate through convolution. Convolution extracts local texture, and ReLU achieves suppression similar to weighted soft thresholding, reducing background noise and enhancing the response to weak targets, improving efficiency and adaptability while preserving the interpretability of sparse modeling.

[0041] WEM module: Taking the current target estimation map as input, it extracts features through convolutional layers and ReLU, stacks multiple sets of convolutional units containing convolution, BatchNorm, and ReLU to integrate local texture and structural features, and outputs a sparse enhancement weight matrix; combined with the local structural difference matrix calculated based on the structural tensor, it constructs an adaptive weight matrix to achieve differentiated weighting of the target region and the background region.

[0042] The method of this invention includes the following steps: establishing a sparse-low-rank decomposition model; deeply unfolding the iterative process into a multi-stage network (including LBRM, STEM, WEM, and introducing DCA and adaptive weighting); end-to-end training (using a combined loss function); and inference segmentation. The network is based on sparse-low-rank decomposition, and iteratively optimizes by using LBRM to model the background, STEM to estimate the target, and WEM to update the weights, combined with DCA and Lagrange multiplier updates, ultimately achieving accurate segmentation of weak targets in complex scenes.

[0043] Example: Figure 1 The figure shows a specific embodiment of an infrared weak target segmentation method based on depth unfolding and adaptive weighting. The steps of this embodiment are as follows:

[0044] S1, take the input infrared image as the observation item and establish a sparse-low-rank decomposition model to express the separation of background and target components;

[0045] S2, the iterative solution process of the sparse-low-rank decomposition is transformed into a trainable multi-stage network through deep expansion. Each stage includes a low-rank background reconstruction module (LBRM), a sparse target estimation module (STEM), and a weight estimation module (WEM); an adaptive weighting matrix W = W is constructed. SD e W SE W SD W is calculated based on the local structural difference of the structure tensor. SE It is generated by a sparse augmentation network and a dual-path complementary attention mechanism is introduced into the LBRM;

[0046] S3, the initial depth unfolding network is trained end-to-end based on the infrared weak target dataset. During the training process, a combination of loss functions of SoftIoU loss, MSE loss and Sparsity loss is used to optimize the network parameters to obtain an optimized infrared weak target segmentation model.

[0047] S4. Use the trained infrared weak target segmentation model to perform inference segmentation on the infrared image to obtain a segmentation result image containing the separated weak targets.

[0048] like Figure 2The diagram shows the neural network model structure of an infrared weak target segmentation method based on deep expansion and adaptive weighting according to the present invention. The model is based on sparse-low-rank matrix factorization and transforms the iterative optimization process into multiple learnable neural network modules through deep expansion. It includes three core sub-modules: LBRM, STEM, and WEM. It introduces DCA attention mechanism and adaptive weighting mechanism and uses a composite loss function for optimization.

[0049] 1. The collected infrared images of weak targets cover a variety of complex scenes, including typical scenes such as cloud cover, ocean wave interference, and building interference; the target size ranges from a few to tens of pixels, ensuring that the dataset can fully verify the model's segmentation ability in weak target and strong interference scenes.

[0050] 2. The sparse-low-rank decomposition model is used to express the separation of background and target components, and is represented as:

[0051] F = T + B

[0052] Where F, T, and B represent the original infrared image, the target, and the background, respectively.

[0053] The constructed adaptive weighting matrix W is expressed as:

[0054] W = W SD ⊙W SE ;

[0055] Among them W SD The calculation of local structural differences based on the structure tensor is expressed by the following formula:

[0056]

[0057] Where h is a hyperparameter used to control the weight sensitivity; d(x,y)=|L1(x,y)-L2(x,y)| represents the difference in structural intensity between the main and secondary directions at pixel (x,y), reflecting the degree of structural change at that point; d max and d min These are the maximum and minimum values ​​of the degree of structural change in the entire image, respectively, used to normalize the structural strength difference.

[0058] W SE Generated by sparse augmentation network This module uses the target estimation map T for the current stage. k Using this as input, rich local texture and structural information is extracted through multi-layer convolution, and the weight distribution is dynamically adjusted to achieve a more accurate response to the target region and background suppression.

[0059] The input image first enters the network initialization module, where the target component, adaptive weight matrix, and Lagrange multipliers are all initialized to zero matrices, serving as the input to the first depth unrolling module. The network undergoes iterative optimization through multiple cascaded depth unrolling modules. Each module outputs updated background B, target T, weights W, and Lagrange multipliers Y, which are then passed to the next module.

[0060] The Lagrange multipliers are dynamically updated based on the current residuals, balancing data consistency constraints to ensure the stability and convergence of the iterative process. The formula can be expressed as: Y k+1 =Y k +(FB k+1 -T k+1 ) / μ k ;

[0061] The degree to which the dynamic equilibrium data consistency constraint F = T + B is satisfied for the update of the Lagrange multiplier Y. The residual term R in the current iteration. k+1 =FB k+1 -T k+1 The observed data F and the decomposition result B are characterized. k+1 +T k+1 The deviation between them. By using the residual R k +1 Multiply by step size factor 1 / μ k , multiplier Y k+1 It is adaptively corrected, thereby gradually reducing constraint bias. This process iteratively optimizes the consistency between variables, while μ... k By controlling the update magnitude, the stability and convergence of the optimization process are ensured, ultimately achieving effective separation of the background and the target.

[0062] 3. The loss function module is used to evaluate the error between the segmentation result and the ground truth label, guiding network optimization. The composite loss function consists of three parts: SoftIoU loss evaluates segmentation accuracy by calculating the overlap between the predicted target and the ground truth label; MSE loss quantifies the difference between the output result and the input image, preserving image details; and SPARSITY loss constrains the sparsity of the target region through the l1 norm, enhancing background suppression. These three factors are weighted to adjust their contribution ratios and jointly optimize the model parameters. The overall loss function formula is:

[0063] L totalLoss =α·SoftIoU_Loss+β·MSE_Loss+γ·SPARSITY_Loss;

[0064] Here, α, β, and γ are hyperparameters used to adjust the relative weights of each loss term during training.

[0065] SoftIoU loss is a loss function based on Intersection over Union (IoU) that evaluates the quality of target region segmentation by measuring the overlap between the predicted segmented region and the ground truth label. The formula for SoftIoU loss is:

[0066]

[0067] Among them, T i This represents the value of the i-th pixel in the actual label, where the pixel value in the target region is 1 and the pixel value in the background region is 0. i This represents the predicted value of the i-th pixel in the segmentation map. N is the total number of pixels in the image.

[0068] To maintain the fidelity of the input image and further enhance the model's generation capability, mean squared error (MSE) loss is added. This loss is used to calculate the error between the model's output and the input image, helping the network retain key information in the image. Its formula is defined as:

[0069]

[0070] Among them, F i This represents the value of the i-th pixel in the input image. This represents the value of the i-th pixel in the model's output image. By minimizing the MSE loss, we ensure that the model can preserve the image's detailed information while optimizing the target segmentation accuracy.

[0071] A sparsity loss function is further introduced. This loss, by applying the L1 norm constraint to the pixel values ​​of the output target image, encourages the model to generate sparse targets, highlighting small target regions and suppressing background interference. The sparsity loss formula is as follows:

[0072]

[0073] Among them, T i This represents the value of the i-th pixel in the target image predicted by the model, where N is the total number of pixels in the image.

[0074] This invention uses two infrared weak target datasets, NUDT-SIRST1K and SIRST-Aug, to verify the model's performance in scenarios with varying complexity. Both datasets have an image resolution of 256×256 pixels, and target regions are labeled with pixel-level annotations.

[0075] The preferred NUDT-SIRST1K dataset contains 1,327 high-resolution infrared images, covering low-altitude flying targets and sea targets under various weather conditions. The preferred SIRST-Aug dataset contains 8,525 training images and 545 test images, covering complex scenes such as aircraft and ships. During training, data augmentation techniques such as random rotation, scaling, contrast adjustment, and Gaussian noise addition were employed to enhance data diversity and model robustness. Experiments were conducted using the PyTorch 1.12 framework on an NVIDIA GeForce RTX 3090 GPU platform equipped with an Intel Xeon Gold 6230R processor, 256GB of RAM, and 24GB of VRAM. The Adam optimizer was used, with a batch size of 16, and a weighted combination of SoftIoU loss, MSE loss, and Sparsity loss was used as the loss function. Peak VRAM usage during training was approximately 18.6GB, with an average epoch time of approximately 128 seconds, and the final model parameter count was 0.681M. After sufficient training, the model achieved an mIoU of 94.37% and an F1-score of 97.10% on the NUDT-SIRST dataset, and an mIoU of 74.13% and an F1-score of 85.15% on the SIRST-Aug dataset. This indicates that the method performs well in target localization, detail preservation, and background suppression, and has high accuracy and good generalization ability.

[0076] like Figure 4 The diagram shows the segmentation results of this invention. The model can effectively suppress interference and accurately segment small targets in complex backgrounds such as clouds and waves, with clear target boundaries and complete details. At the same time, the lightweight design achieves a balance between high precision and high efficiency, making it suitable for resource-constrained scenarios such as nighttime monitoring and maritime rescue, providing strong technical support for the accurate detection of small infrared targets.

[0077] In summary, this invention models the segmentation of weak infrared targets as a sparse-low-rank matrix factorization problem. The backbone network adopts a deep unfolded module stacking structure. The core modules include a low-rank background reconstruction module (LBRM), a sparse target estimation module (STEM), and a weight estimation module (WEM). A dual-path complementary attention (DCA) mechanism is introduced, and a combined loss function consisting of SoftIoU loss, MSE loss, and Sparsity loss is used.

[0078] The model was trained and validated on two publicly available infrared small target datasets, NUDT-SIRST and SIRST-Aug. The NUDT-SIRST dataset contains a variety of typical complex backgrounds, including clouds, sea surfaces, buildings, and other infrared scenes. The target scale is mostly a few to tens of pixels, and the signal-to-noise ratio is less than 3dB, reflecting the difficult conditions in the detection of weak targets in reality. The SIRST-Aug dataset has been enhanced in various ways to cover more complex background combinations and weak signal scenes, thus ensuring that the model has good generalization performance under various backgrounds, sizes, and contrasts.

[0079] The infrared image is modeled as a superposition of a low-rank background B and sparse targets T. The designed depth unfolding network consists of multiple cascaded depth unfolding modules, each corresponding to a stage of optimization iteration. When the network is input, the target estimate T is... 0 Adaptive weight matrix W 0 and Lagrange multipliers Y 0 The matrix is ​​initialized to zero and used as the initial state input to the first module. During iteration, each module receives the outputs T, Y, and W from the previous stage and processes them sequentially through the Low-Rank Background Reconstruction (LBRM) module, the Sparse Target Estimation (STEM) module, and the Weight Estimation (WEM) module, achieving progressive optimization from initial separation to precise segmentation. LBRM estimates the low-rank background, STEM extracts sparse targets, and WEM generates a new adaptive weight matrix, which is then passed to the next stage for iterative updates. Finally, the optimized B, T, W, and Y are output at the end of the network.

[0080] In the low-rank background reconstruction module, a dual-path complementary attention DCA mechanism is introduced. Local detail features and global context features are modeled through channel and spatial paths, respectively. The former enhances the response of target edges and details, while the latter supplements long-range dependencies and semantic information. After the features of the two paths are fused, the network's response strength to weak targets can be significantly improved, while suppressing false detections caused by structured backgrounds, thereby improving the segmentation accuracy and robustness in complex backgrounds.

[0081] A composite loss function is employed, consisting of a weighted average of SoftIoU loss, MSE loss, and Sparsity loss. SoftIoU optimizes the overlap between the predicted target and the ground truth label, improving segmentation accuracy; MSE loss maintains consistency between the reconstructed background and the original image, avoiding excessive weakening of background information; and Sparsity loss, based on the l1 norm, constrains the sparsity of the target matrix, effectively suppressing noise responses. During training, the same hyperparameters and iterative strategies are used on both datasets to ensure the stability of the optimization process and the comparability of the results. Through this design, this invention significantly improves inference efficiency and interpretability while maintaining high detection accuracy, making it suitable for infrared weak target segmentation tasks in various complex backgrounds.

[0082] Example Conclusion:

[0083] By introducing a lightweight feature extraction mechanism, the feature extraction process for infrared weak target segmentation is optimized, maintaining efficient feature representation capabilities while reducing computational complexity. The designed deep unfolding modules LBRM, STEM, and WEM not only enhance target-background separation capabilities by fusing sparse-low-rank decomposition priors, but also simplify the network structure through modular design, further reducing the number of model parameters to only 0.681B, thus significantly improving computational efficiency and resource utilization. Furthermore, by combining a dual-path complementary attention (DCA) module, the segmentation accuracy of weak targets in complex backgrounds such as clouds and buildings is significantly improved by capturing local details and global context in parallel, enhancing the model's robustness to low-contrast and high-noise scenes. By introducing an adaptive weighting mechanism composed of a local structure difference matrix and a sparse enhancement matrix, and a composite loss function fusing SoftIoU, MSE, and Sparsity, this invention not only improves the prediction accuracy of target boundaries but also enhances the ability to identify weak targets. Experimental results show that the present invention performs excellently on the NUDT-SIRST and SIRST-Aug datasets, with mIoU of 94.37% and 74.13% respectively, and F1-scores of 97.10% and 85.15% respectively. It has higher segmentation accuracy and recall rate, and significantly reduces computational resource consumption. It is especially suitable for resource-constrained practical scenarios such as night monitoring and maritime rescue, and can provide strong technical support for high-precision detection of weak infrared targets, promoting the intelligent development of related fields.

[0084] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for infrared dim small target segmentation based on deep unfolding and adaptive weighting, characterized in that The method comprises the following steps: S1, taking an input infrared image as an observation item and establishing a sparse-low-rank decomposition model to express the separation of background and target components; S2, the iterative solution process of the sparse and low-rank decomposition is transformed into a trainable multi-stage network by deep unfolding, each stage including a low-rank background reconstruction module (LBRM), a sparse target estimation module (STEM) and a weight estimation module (WEM); wherein the weight estimation module (WEM) generates sparse enhancement weights W SE , combined with the local structure difference weights W SD calculated by the structure tensor, to obtain an adaptive weighting matrix W = W SD ⊙ W SE , which is used to constrain the sparse estimation of the sparse target estimation module (STEM) and combine the double-path complementary attention mechanism in the low-rank background reconstruction module (LBRM) to enhance the separation of the target and the background; S3, performing end-to-end training on the initial deep unfolding network based on the infrared dim target data set, adopting a combined loss function composed of a SoftIoU loss, an MSE loss and a Sparsity loss to optimize network parameters during the training process, and realizing the synchronous optimization of target precision, structure fidelity and sparse characteristics; Wherein: the SoftIoU loss is used to improve the segmentation overlap precision of small target regions, the MSE loss is used to maintain the consistency of input and output images to avoid detail loss, and the Sparsity loss is used to strengthen the target sparsity and suppress background noise interference; through the joint constraint of the three, the synchronous optimization of target precision, structure fidelity and sparse characteristics is realized, so as to ensure the complete reservation and high-precision segmentation of dim targets in complex background. S4, using the trained infrared dim target segmentation model to perform inference segmentation on the infrared image to obtain a segmentation result image containing separated dim targets.

2. The infrared dim small target segmentation method based on depth unfolding and adaptive weighting according to claim 1, characterized in that: Each stage of the deep unfolding network comprises a low-rank background reconstruction module (LBRM), a sparse target estimation module (STEM) and a weight estimation module (WEM), wherein each module is respectively used for low-rank background modeling, sparse target estimation and adaptive weight updating to realize the separation of background component B and sparse target component T in the observation image F=T+B.

3. The infrared dim small target segmentation method based on depth unfolding and adaptive weighting according to claim 1 or 2, characterized in that: The low-rank background reconstruction module (LBRM) replaces the traditional optimization process through an end-to-end learnable manner, including the steps of convolution layer mapping, double-path parallel attention module (DCA) introduction, batch normalization and ReLU transformation, convolution stacking unit feature extraction and convolution channel compression output background estimation image; specifically comprising the following steps: first, the input is mapped to the feature space by using the convolution layer; then the double-path parallel attention module (DCA) is introduced; then the nonlinear transformation is realized through batch normalization and ReLU; then the deep features are extracted through the stacking unit of convolution-normalization-activation; finally, the background estimation image is output through convolution compression channel.

4. The infrared dim small target segmentation method based on depth unfolding and adaptive weighting according to claim 3, characterized in that: The double-path parallel attention module (DCA) comprises: the first path combines channel attention and spatial attention to highlight the target region; the second path performs global context modeling based on the Query-Key-Value mechanism, calculates the importance of space and channel, and integrates long-distance dependency information; the two paths are calculated in parallel and then fused and output to enhance the precision and robustness of background modeling.

5. The infrared dim small target segmentation method based on depth unfolding and adaptive weighting according to claim 1 or 2 or 4, characterized in that: The low-rank background reconstruction module (LBRM) is a learnable subnetwork; in structure, it is first subjected to 3*3 convolution and ReLU initial feature extraction, then fuses local spatial background information through multiple 3*3 convolution blocks, and finally outputs the reconstructed background features by adjusting the channel number through convolution. Specifically, the low-rank background reconstruction module (LBRM) first applies a 3x3 convolution and ReLU activation to the input features for initial feature extraction, then captures and fuses local spatial background information through a plurality of consecutive 3x3 convolution blocks, and finally adjusts the number of channels to the input channel number through a layer of convolution to output the reconstructed background features. The convolution operation is used to extract the local texture and structure information of the background, and the ReLU activation produces a nonlinear inhibition effect, thereby effectively suppressing target interference and noise.

6. The infrared dim small target segmentation method based on depth unfolding and adaptive weighting according to claim 1, characterized in that: The weight estimation module (WEM) comprises an input convolutional layer for extracting initial features, a plurality of convolution-normalization-activation units for multi-scale texture feature extraction and non-linear mapping, and an output mapping layer for transforming the features into sparse enhancement weights W consistent with the size of the input image SE The module can adaptively adjust the weights according to the sparse target distribution at different stages to dynamically optimize the saliency of the target region in the iterative updating process.

7. The infrared dim small target segmentation method based on depth unfolding and adaptive weighting according to claim 1 or 2, characterized in that: In the combined loss function, the SoftIoU loss is used to optimize the segmentation overlap accuracy, the weight estimation module (WEM) loss maintains the image details, and the Sparsity loss enhances the target sparsity. The total loss function is: L totalLoss = a * SoftlOu_Loss + b * MSE_Loss + g * SPARSITY_Loss; wherein α, β, and γ are hyperparameters used to adjust the relative weights of each loss term in the training process.

8. The infrared dim small target segmentation method based on depth unfolding and adaptive weighting according to claim 1, characterized in that: The infrared dim small target dataset is NUDT-SIRST and SIRST-Aug, the former contains multiple complex backgrounds and has low target signal-to-noise ratio, and the latter is an enhanced version to improve the model generalization.

9. The infrared dim small target segmentation method based on depth unfolding and adaptive weighting according to claim 1, characterized in that: The W SD The calculation formula is: where h is a weight sensitivity hyper-parameter, d(x, y) represents the pixel point structure intensity difference; specifically: d(x, y) = |L1(x, y) - L2(x, y)| represents the difference between the structure intensity of the main direction and the auxiliary direction at the (x, y) pixel point.

10. The infrared dim small target segmentation method based on depth unfolding and adaptive weighting according to claim 1, characterized in that: The Lagrange multiplier Y update formula is also included: Y k+1 = Y k + (F - B k+1 - T k+1 ) / μ k , to balance the data consistency constraint; wherein Y k is the Lagrange multiplier at the kth iteration, used to balance the data consistency constraint; Y k+1 is the updated Lagrange multiplier at the (k+1)th iteration; F is the original infrared image; B k+1 is the low-rank background estimation result at the (k+1)th iteration; T k+1 is the sparse target estimation result at the (k+1)th iteration; μ k is the step coefficient at the kth iteration, used to control the updating amplitude of the Lagrange multiplier, guaranteeing the stability and convergence of the optimization process.

Images