Camouflage target segmentation system and method based on multi-scale boundary fusion

Abstract

The invention discloses a camouflage target segmentation system based on multi-scale boundary fusion. The camouflage target segmentation system comprises a data division module for acquiring a training data set; the feature extraction module performs feature extraction based on a backbone network to obtain multi-level hierarchical feature information; the edge priori module combines the low-level features and the high-level features, performs enhancement by using a parameter-free attention mechanism, and obtains camouflage target edge information; the multi-scale edge enhancement module extracts multi-scale edge feature information by using global average pooling processing, and obtains output features in combination with camouflage target edge information; the multi-scale detail fusion module uses a Hadamard product to be combined with the output features to obtain enhanced information; and the up-sampling module combines the enhanced information with low-level features based on PFNet, and takes a prediction result of the last-level focusing module as a final segmentation result. According to the method, the defects of an existing camouflage target detection technology in the aspects of boundary detail capture and multi-scale feature fusion are overcome, and the precision and robustness of camouflage target detection are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and in particular to a camouflaged target segmentation system and method based on multi-scale boundary fusion. Background Art

[0002] Camouflaged object detection (COD) is an important research area in computer vision. It aims to detect objects that closely resemble their surroundings within complex backgrounds. Due to their ability to seamlessly blend into their surroundings, COD has demonstrated widespread application value and importance across multiple fields.

[0003] Existing camouflaged object detection methods can be broadly categorized into two types: network architecture and learning paradigm. Network architecture describes how the input and output configurations within a model are structured, such as linear, aggregate, and branched network architectures, effectively exploring discriminative camouflaged object features for COD. Learning paradigm refers to the methods used for model learning and adaptation. Despite progress, existing methods still suffer from the following key drawbacks: Due to the high spatial similarity between camouflaged objects and their backgrounds, these methods are susceptible to interference in complex backgrounds, making it difficult to effectively distinguish between them. Most existing methods focus solely on high- or low-frequency features, neglecting important information in the mid-frequency bands and resulting in incomplete utilization of frequency domain information. Furthermore, frequency features are susceptible to background noise, resulting in inaccurate feature representation. While performance has been improved through multi-scale feature fusion and multi-task learning, existing models still struggle to process object boundary details. This is particularly true for complex backgrounds and multi-scale object detection, where detailed object information is often overlooked, resulting in reduced segmentation accuracy.

[0004] Therefore, inventing a camouflaged target segmentation system based on multi-scale boundary fusion, which can accurately capture the target edge information, highlight the key boundaries of the target, improve the model's sensitivity to the target boundary, retain detailed features while enhancing semantic information, and further improve the accuracy and robustness of camouflaged target detection has become a technical problem that needs to be solved urgently. Summary of the Invention

[0005] The purpose of the present invention is to provide a camouflaged target segmentation system based on multi-scale boundary fusion, which can solve the shortcomings of existing methods in capturing boundary details and fusion of multi-scale features, and significantly improve the performance of camouflaged target detection.

[0006] To achieve this goal, the present invention designs a camouflaged target segmentation system based on multi-scale boundary fusion, which includes:

[0007] The data partitioning module is used to divide the disguised target image data set into the training data set;

[0008] The feature extraction module uses any image in the training data set as a feature map and extracts features from the feature map through the backbone network to obtain multi-level hierarchical feature information. The multi-level hierarchical feature information contains low-level features containing detail information and high-level features containing semantic information.

[0009] The edge prior module is used to combine low-level features and high-level features, and use a parameter-free attention mechanism to enhance the disguised target area information in the feature map, and extract the disguised target edge information based on the enhanced disguised target area information;

[0010] The multi-scale edge enhancement module uses global average pooling to process feature information at all levels, extracts multi-scale edge feature information, combines it with the camouflaged target edge information, and enhances the details of the multi-scale edge feature information through fine-grained detail feature enhancement to obtain enhanced detail features. The enhanced detail features are combined with the camouflaged target edge information to obtain the enhanced output features of each level;

[0011] The multi-scale detail fusion module uses the Hadamard product (a matrix operation based on element-wise multiplication) to combine the enhanced output features of all levels to obtain the enhanced multi-scale camouflaged target edge information;

[0012] The upsampling module uses the positioning module and multi-level focusing module in the PFNet (Positioning Focus Network camouflaged target detection model) to combine the enhanced multi-scale camouflaged target edge information with low-level features, and passes the combined information in the multi-level focusing module in sequence. The prediction result of the last-level focusing module is used as the final segmentation result.

[0013] Preferably, it further includes a data evaluation module for dividing the disguised target image data set into a test data set, and calculating an evaluation index based on the test data set for evaluating the final segmentation result.

[0014] Beneficial effects of the present invention: The present invention proposes a camouflaged target segmentation system based on multi-scale boundary fusion, which combines low-level features and high-level features through an edge prior module, and utilizes a parameter-free attention mechanism SimAM to adaptively enhance key area information in the feature map, thereby accurately extracting edge information closely related to the target, which can effectively solve the problem of difficulty in accurately distinguishing between the target and the background in the existing technology, and significantly improve the segmentation and positioning performance of target detection; through multi-scale edge feature extraction and detail enhancement, the key boundaries of the target are highlighted, and the edge information is integrated into the feature learning process, which effectively solves the problem that the existing technology easily ignores target details in complex backgrounds and multi-scale target detection, and improves the model's sensitivity to target boundaries, thereby improving segmentation accuracy; through the Hadamard product operation, multiple layers of features are integrated to enhance semantic information while retaining detail features, significantly improving the accuracy and robustness of camouflaged target detection, so that the model can generate more accurate and detail-rich segmentation results when processing complex scenes. The present invention effectively solves the shortcomings of existing camouflaged target detection technology in boundary detail capture and multi-scale feature fusion through multi-scale boundary fusion technology and modular design, significantly improves the performance of camouflaged target detection, and has broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a structural schematic diagram of the present invention;

[0016] Figure 2 It is a network process framework diagram;

[0017] Figure 3 This is a schematic diagram of the SEP module framework;

[0018] Figure 4 This is a schematic diagram of the MEE module framework;

[0019] Figure 5 A comparison chart of experimental results. DETAILED DESCRIPTION

[0020] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings of the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0021] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0022] Example 1

[0023] A camouflaged target segmentation system based on multi-scale boundary fusion, such as Figure 1 As shown, it includes:

[0024] The data partitioning module is used to divide the disguised target image data set into the training data set;

[0025] The feature extraction module uses any image in the training data set as a feature map and extracts features from the feature map through the backbone network to obtain multi-level hierarchical feature information. The multi-level hierarchical feature information contains low-level features containing detail information and high-level features containing semantic information.

[0026] The edge prior module is used to combine low-level features and high-level features, and use a parameter-free attention mechanism to enhance the disguised target area information in the feature map, and extract the disguised target edge information based on the enhanced disguised target area information;

[0027] The multi-scale edge enhancement module uses global average pooling to process feature information at all levels to extract multi-scale edge feature information. This multi-scale edge feature information is combined with the camouflaged target edge information and enhanced through fine-grained detail feature enhancement to obtain enhanced detail features. This enhanced detail feature is then combined with the camouflaged target edge information to obtain the enhanced output features at each level.

[0028] The multi-scale detail fusion module uses Hadamard product to combine the enhanced output features of all levels to obtain the enhanced multi-scale camouflaged target edge information;

[0029] The upsampling module uses the positioning module and multi-level focusing module in PFNet to combine the enhanced multi-scale camouflaged target edge information with low-level features, and passes the combined information in the multi-level focusing modules in sequence. The prediction result of the last level focusing module is used as the final segmentation result.

[0030] In the above technical solution, to address the problem that it is difficult to accurately identify camouflaged objects with complete and detailed object structures in the existing technology, and the detection results are not accurate enough, the present invention proposes a camouflaged target segmentation network based on multi-scale boundary fusion (Multiscale Edge Semantics and Detail Infusion Network, EFNet): The network proposes three innovative modules: the SimAM-based Edge Prior Module (SEP) based on parameter-free attention mines edge features related to the object from low-level features containing local edge details and high-level features containing global position information; the Multiscale Edge-Enhanced Module (MEE) uses the edge enhancer to enhance edge perception at multiple scales on each layer of feature maps extracted by the encoder to highlight the key boundaries of the target, and then integrates them with the edge features from the edge prior module; the Multiscale Semantic Detail Fusion Module (MDSF) integrates each layer of feature maps, i.e., low-level features containing local edge details and high-level features containing global position information, through the Hadamard product operation, thereby enhancing the semantics and details of the features.

[0031] In the above technical solution, in the EFNet (Edge-Fusion Network) of the present invention, the PM and FM modules of PFNet are retained and embedded in the upsampling stage, but their input features are replaced by enhanced multi-scale edge features (from the MDSF module). By combining PFNet with the edge enhancement module, EFNet achieves higher boundary accuracy and structural integrity in camouflaged target detection in complex scenes, while inheriting the high efficiency of PFNet.

[0032] In the above technical solution, a camouflaged target segmentation system based on multi-scale boundary fusion is defined, including six modules: data partitioning, feature extraction, edge prior, multi-scale edge enhancement, multi-scale detail fusion, and upsampling. This protects the overall architecture of the system and ensures the innovative combination of modular design (SEP, MEE, MDSF) and PFNet upsampling.

[0033] In the above technical solution, the camouflaged target image dataset includes the CHAMELEON dataset, the CAMO dataset and the COD10K dataset, wherein the CHAMELEON dataset has 76 images, all of which are used as test datasets; the CAMO dataset has 2,500 images, of which 80% of the image data is used as training datasets and 20% of the image data is used as test datasets; COD10K is currently the largest camouflaged target detection dataset, of which 60% of the image data is used as training datasets and 40% of the image data is used as test datasets, and the test dataset and training dataset are divided by the Baseline-PFNet deep learning model.

[0034] In the above technical solution, by specifying the data set division method (CHAMELEON, CAMO, COD10K), clarifying the training and testing ratios, and limiting the specific application of the data set, the scientific nature of model training and evaluation is ensured.

[0035] The above technical solution also includes a data evaluation module for dividing the disguised target image data set into a test data set, calculating an evaluation index based on the test data set, and evaluating the final segmentation result.

[0036] In the above technical solution, the evaluation indicators include E (edge segmentation accuracy) and S (structural integrity). Edge segmentation accuracy is used to measure the algorithm's ability to identify target edge details, especially the ability to accurately segment the fuzzy boundary between the camouflaged target and the background. Structural integrity is used to measure the integrity and coherence of the target's overall structure in the segmentation result, avoiding segmentation breaks or morphological distortion caused by the loss of local features. In this evaluation, four evaluation indicators are mainly used, namely S-measure (Sα), F-measure (Fβ), Mean Absolute Error (MAE) and E-measure (Eφ), among which:

[0037] S-measure (Sα) predicts the spatial structure similarity between the result image and the GT image, which combines object perception (So) and region perception (Sr) evaluation;

[0038] The TF-measure value (Fβ) is used to calculate the relationship between precision (P) and recall (R);

[0039] E-measure (Eφ) evaluates the local and global similarity between the result image and the GT image;

[0040] Mean Absolute Error (MAE) is the average absolute difference per pixel between the normalized prediction result image and the GT image.

[0041] In the above technical solution, the evaluation results can be fed back into the model optimization to improve iteration efficiency.

[0042] In the above technical solution, the evaluation indicators are calculated through the test set, emphasizing the integrity of the system performance verification.

[0043] In the above technical solution, the feature extraction module uses the PVT-B2 (PVTv2: Improved Baselines with Pyramid Vision Transformer backbone network to extract features and obtain 5-level hierarchical feature information. Each level of feature information is represented by f i , where i∈{1,2,3,4,5}, outputs high-level features f5 containing semantic information and low-level features f2 containing detail information.

[0044] In the above technical solution, the present invention uses PVT-B2 as the feature encoder for feature extraction because PVTv2 has achieved remarkable performance in classification, detection and segmentation. This model is not the only one that can achieve this. There are many backbone networks that can perform feature extraction, such as ResNet and EfficientNet. There are no specific requirements.

[0045] In the above technical solution, low-level features come from shallow networks, are rich in spatial information, and the feature resolution of spatial information is relatively high; high-level features come from deep networks, are rich in semantic information, and the feature resolution of semantic information is relatively low.

[0046] In the above technical solution, by specifying the feature extraction module to use the PVT-B2 backbone network and output 5-level features, PVT-B2 is used to balance computational efficiency and feature expression capabilities, thereby protecting the specific implementation method of feature extraction, clarifying the roles of low-level features and high-level features, and optimizing the input of subsequent modules.

[0047] In the above technical solution, if Figure 3 As shown in Figure 2, the specific operation steps of the edge prior module are as follows:

[0048] Two 1×1 convolutions are used for feature channel processing, and the number of channels of the low-level feature f2 is adjusted to 64, and the number of channels of the high-level feature f5 is adjusted to 256; then the parameter-free SimAM attention is used to process the low-level features and high-level features separately to focus on and segment the camouflaged target; the SimAM module does not require additional parameters and can adaptively enhance the feature information of important areas in the feature map. Then, the low-level features f2 and high-level features f5 processed separately by SimAM are connected in series to achieve cross-scale feature fusion, combining detail information and semantic information to generate a fused feature; two 3×3 convolutions and one 1×1 convolution are used to restore the number of channels and optimize the feature information; SimAM attention is applied again to focus on the fused features to obtain the camouflaged target edge information f e .

[0049] In the above technical solution, by refining the operation steps of the edge prior module (SEP), the SimAM attention mechanism is used to reduce the computational complexity and adaptively focus on key areas; low-level details and high-level semantics are combined to improve the accuracy of edge feature extraction.

[0050] In the above technical solution, if Figure 4 As shown in Figure 2, the specific way in which the detail features are enhanced by the multi-scale edge enhancement module is as follows:

[0051] The average pooling and 1×1 convolution strategy is adopted to obtain the input feature f i Extract multi-scale information from i1 、f i2 、f i3 、f i4 , use edge enhancer to enhance information f i2 、f i3 、f i4 Edge perception, to obtain enhanced multi-scale information Ensure that the key boundaries of the object can be highlighted, fuse the extracted multi-scale edge information to improve the precision of edge prediction, and obtain the enhanced features of each level The specific calculation formula is:

[0052] f i1 =Conv 1×1 (f i )

[0053] f i2 =AP(Conv 1×1 (f i1 ))

[0054] f i3 =AP(Conv 1×1 (f i2 ))

[0055] f i4 =AP(Conv 1×1 (f i3 ))

[0056]

[0057] Among them, Conv represents the convolution operation, EE represents the edge enhancer strengthening processing, AP represents the average pooling, and concat represents the operation of fusing the extracted multi-scale edge information.

[0058] In the above technical solution, the specific calculation method of the multi-scale edge enhancement module (MEE) is specified, such as extracting edge information through different pooling scales, enhancing the sensitivity of complex boundaries, optimizing feature expression through the edge enhancer (EE), reducing background interference, and protecting the multi-scale processing flow of the MEE module.

[0059] In the above technical solution, f i1 By input feature f i It is directly generated by 1×1 convolution without pooling operation, which preserves the spatial resolution of the original feature map and has the richest detail information.

[0060] f i2 By f i1 After 1×1 convolution and average pooling (AP), average pooling will reduce the resolution of the feature map and reduce the spatial size;

[0061] f i3 and f i4 It is gradually generated by repeatedly passing through 1×1 convolution and average pooling (AP). Each level of pooling further reduces the resolution, expands the receptive field, enhances the semantic information, but gradually reduces the details.

[0062] In the above technical solution, if Figure 4 As shown in Figure 2, the enhanced detail features are combined with the camouflaged target edge information to obtain the specific calculation formula for the enhanced output features of each level:

[0063]

[0064] Among them, the enhanced features of each level and the downsampled camouflaged target edge information f e Perform element-wise multiplication and combine with the enhanced features at each level Adding them together, we get the initial fusion features. In order to further optimize the fusion features, we use global average pooling to aggregate the initial fusion features. Then, we use 1×1 convolution and Sigmoid function to get the feature information of each level f iThe corresponding channel weights are multiplied by the initial fusion features to obtain the output feature f oi .

[0065] In the above technical solution, downsampling is a commonly used image processing technology used to reduce the spatial resolution of an image, thereby reducing the amount of calculation and memory usage.

[0066] In the above technical solution, by defining the mathematical formula for the fusion of features and edge information in the MEE module, for example, by generating channel weights through the Sigmoid function and global pooling to aggregate global information, the importance of features is adaptively adjusted to improve the structural coherence of the segmentation results.

[0067] In the above technical solution, the specific operation steps of the multi-scale camouflaged target edge information after the multi-scale detail fusion module is enhanced are:

[0068] The multi-scale detail fusion module combines the feature information of each level f i And the camouflaged target edge information f e Input into the multi-scale edge enhancement module to obtain the output feature f of each layer oi ;

[0069]

[0070] Among them, Up represents upsampling, Down represents downsampling, and by selecting the output feature f of the second layer o2 As the target reference feature, it is used to adjust the resolution of feature maps of other layers to make them consistent with f o2 The resolution of the feature f′ is matched to obtain the adjusted feature f′ oi After adjusting the feature map resolution, a 3×3 convolution smoothing process is applied to each feature map. Finally, the adjusted feature map is combined with the enhanced output features of all levels through the Hadamard product to enhance the semantic information and detail information of the output features. Figure 5 As shown, the enhanced multi-scale camouflaged target edge information f is obtained out , the specific calculation formula of the enhanced multi-scale camouflaged target edge information is:

[0071] f out =Hadamard(f′ o1 ,f′ o2 ,f′ o3 ,f′ o4 ,f′ o5 ).

[0072] In the above technical solution, the enhanced output features are combined with the specified low-level feature information to achieve complementarity and effectively balance semantic information and detail information, significantly improving the richness and accuracy of feature representation and providing more reliable feature support for subsequent segmentation tasks.

[0073] In the above technical solution, the feature scale is unified through upsampling and downsampling to avoid information loss, and the high-level and low-level features are combined through the Hadamard product to retain details while enhancing semantic expression.

[0074] In the above technical solution, the model proposed by the present invention does not change the key modules of PFNet (positioning module PM and multi-level focusing module FM), but proposes three additional modules (edge prior module SEP, multi-scale detail fusion module MDSF, multi-scale edge enhancement module MEE). EFNet has achieved significant improvements in all four indicators; in E, the network of the present invention improved by 1.9%, 1.1%, and 1.4% on the CHAMELEON, CAMO, and COD-10K datasets, respectively. This can be attributed to SEP and MEE. The SEP module deeply mines edge features by fusing high-level semantic features with low-level features, and the MEE module focuses on refining multi-scale edge information and fusing it, thereby improving edge segmentation accuracy. The network proposed by the present invention is 2.4%, 1.2%, and 0.9% higher than PFNet in terms of S, which shows that the model of the present invention performs well in mining the complete structure of the camouflaged object, which can be attributed to the integration of multi-level features by the MDSF module.

[0075] Example 2

[0076] A camouflaged target segmentation algorithm based on multi-scale boundary detail fusion, such as Figure 2 As shown in the figure, the disguised image dataset is divided into a training dataset and a test dataset; the input image I is extracted through the PVT-B2 backbone network, and these features are then reduced through five convolutional layers to obtain multi-level feature information f i , i∈{1,2,3,4,5}, the fifth layer features are fused with feature information of all levels to form a high-level feature f5 containing global position information, combined with low-level features f2 and high-level features f5, and the parameter-free attention mechanism SimAM is used to adaptively enhance the key area information in the feature map, thereby extracting the edge information f that is closely related to the target e , for each level of feature f i Multi-scale edge feature extraction and detail enhancement are performed, and these enhanced features are combined with edge information f e Combined, the enhanced feature f is finally obtained oi , integrating all enhanced features f oi , and the output f is obtainedout ,In the upsampling stage, the positioning module and focusing module based on PFNet are used, f out It is added with the low-level feature f2 and the prediction of the last FM module is used as the final segmentation result in the testing phase.

[0077] A camouflaged target segmentation method based on multi-scale boundary fusion includes the following steps:

[0078] Divide the camouflaged target image dataset into a training dataset;

[0079] Using any image in the training data set as a feature map, the backbone network extracts features from the feature map to obtain multi-level hierarchical feature information, which contains low-level features containing detail information and high-level features containing semantic information;

[0080] Combining low-level features and high-level features, and utilizing a parameter-free attention mechanism to enhance the camouflaged target region information in the feature map, and extracting the camouflaged target edge information based on the enhanced camouflaged target region information;

[0081] Global average pooling is used on all levels of feature information to extract multi-scale edge feature information. This information is combined with the camouflaged target edge information and enhanced through fine-grained detail feature enhancement to obtain enhanced detail features. This enhanced detail feature is then combined with the camouflaged target edge information to obtain the enhanced output features of each level.

[0082] The Hadamard product is used to combine the enhanced output features of all levels to obtain the enhanced multi-scale camouflaged target edge information;

[0083] Using the positioning module and multi-level focusing module in PFNet, the enhanced multi-scale camouflaged target edge information is combined with the low-level features, and the combined information is passed sequentially in the multi-level focusing modules. The prediction result of the last level focusing module is used as the final segmentation result.

[0084] Example 3

[0085] A computer program product includes a computer program, which implements the steps of the method described in Example 2 when executed by a processor.

[0086] The contents not described in detail in this specification belong to the prior art known to professional and technical personnel in this field.

[0087] It will be understood by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0088] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A system that specifies the functions of a box or boxes.

[0089] These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture including an instruction system that is implemented in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0090] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0091] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit its scope of protection. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that after reading the present invention, those skilled in the art may still make various changes, modifications or equivalent substitutions to the specific implementation methods of the invention, but these changes, modifications or equivalent substitutions are all within the scope of protection of the pending claims of the invention.

Claims

1. A camouflaged target segmentation system based on multi-scale boundary fusion, characterized in that: include: The data partitioning module is used to divide the disguised target image data set into the training data set; The feature extraction module uses any image in the training data set as a feature map and extracts features from the feature map through the backbone network to obtain multi-level hierarchical feature information. The multi-level hierarchical feature information contains low-level features containing detail information and high-level features containing semantic information. The edge prior module is used to combine low-level features and high-level features, and use a parameter-free attention mechanism to enhance the disguised target area information in the feature map, and extract the disguised target edge information based on the enhanced disguised target area information; The multi-scale edge enhancement module uses global average pooling to process feature information at all levels, extracts multi-scale edge feature information, combines it with the camouflaged target edge information, and enhances the details of the multi-scale edge feature information through fine-grained detail feature enhancement to obtain enhanced detail features. The enhanced detail features are combined with the camouflaged target edge information to obtain the enhanced output features of each level; The multi-scale detail fusion module uses Hadamard product to combine the enhanced output features of all levels to obtain the enhanced multi-scale camouflaged target edge information; The upsampling module uses the positioning module and multi-level focusing module in PFNet to combine the enhanced multi-scale camouflaged target edge information with low-level features, and passes the combined information in the multi-level focusing modules in sequence. The prediction result of the last level focusing module is used as the final segmentation result.

2. The camouflaged target segmentation system based on multi-scale boundary fusion according to claim 1, characterized in that: The camouflaged target image dataset includes the CHAMELEON dataset, the CAMO dataset, and the COD10K dataset, wherein the image data of the CHAMELEON dataset are all used as the test dataset; 80% of the image data in the CAMO dataset are used as the training dataset, and 20% of the image data are used as the test dataset; 60% of the image data in the COD10K dataset are used as the training dataset, and 40% of the image data are used as the test dataset, and the test dataset and the training dataset are divided by the Baseline-PFNet deep learning model.

3. The camouflaged target segmentation system based on multi-scale boundary fusion according to claim 1, characterized in that: It also includes a data evaluation module for dividing the disguised target image dataset into a test dataset, calculating evaluation indicators based on the test dataset, and evaluating the final segmentation results.

4. The camouflaged target segmentation system based on multi-scale boundary fusion according to claim 1, characterized in that: In the feature extraction module, the PVT-B2 backbone network is used to extract features and obtain 5-level hierarchical feature information. Each level of feature information is represented by f i , where i∈{1,2,3,4,5}, outputs high-level features f5 containing semantic information and low-level features f2 containing detail information.

5. The camouflaged target segmentation system based on multi-scale boundary fusion according to claim 4, characterized in that: The specific operation steps of the edge prior module are: Two 1×1 convolutions are used to process feature channels, adjusting the number of channels of the low-level feature f2 to 64 and the number of channels of the high-level feature f5 to 256. The parameter-free SimAM attention is then used to process the low-level and high-level features separately to focus on segmenting the camouflaged target. The SimAM module adaptively enhances the feature information in the feature map, and then concatenates the low-level features f2 and high-level features f5 processed by SimAM to achieve cross-scale feature fusion, combining detail information and semantic information to generate a fused feature. The number of channels is restored and the feature information is optimized through two 3×3 convolutions and one 1×1 convolution; SimAM attention is applied again to focus on the fused features to obtain the camouflaged target edge information f e .

6. The camouflaged target segmentation system based on multi-scale boundary fusion according to claim 5, characterized in that: The specific way in which the detailed features after the multi-scale edge enhancement module are enhanced is as follows: The average pooling and 1×1 convolution strategy is adopted to obtain the input feature f i Extract multi-scale information from i1 、f i2 、f i3 、f i4 , use edge enhancer to enhance information f i2 、f i3 、f i4 Edge perception, to obtain enhanced multi-scale information The extracted multi-scale edge information is fused to obtain the enhanced features of each level The specific calculation formula is: f i1 =Conv 1×1 (f i ) f i2 =AP(Conv 1×1 (f i1 )) f i3 =AP(Conv 1×1 (f i2 )) f i4 =AP(Conv 1×1 (f i3 )) Among them, Conv represents the convolution operation, EE represents the edge enhancer strengthening processing, AP represents the average pooling, and concat represents the operation of fusing the extracted multi-scale edge information.

7. The camouflaged target segmentation system based on multi-scale boundary fusion according to claim 6, characterized in that: The enhanced detail features are combined with the camouflaged target edge information to obtain the specific calculation formula for the enhanced output features of each level: The formula means that the features after each level enhancement and the downsampled camouflaged target edge information f e Perform element-wise multiplication and combine with the enhanced features at each level Add together to get the initial fusion features, and use global average pooling to aggregate the initial fusion features; Then, 1×1 convolution and Sigmoid function are used to obtain the feature information f of each level. i The corresponding channel weights are multiplied by the initial fusion features to obtain the output feature f oi .

8. The camouflaged target segmentation system based on multi-scale boundary fusion according to claim 7, characterized in that: The specific operation steps of the multi-scale camouflaged target edge information after the multi-scale detail fusion module is enhanced are: The multi-scale detail fusion module combines the feature information of each level f i And the camouflaged target edge information f e Input into the multi-scale edge enhancement module to obtain the output feature f of each layer oi ; Among them, Up represents upsampling, Down represents downsampling, and by selecting the output feature f of the second layer o2 As the target reference feature, it is used to adjust the resolution of feature maps of other layers to make them consistent with f o2 The resolution of the feature f′ is matched to obtain the adjusted feature f′ oi After adjusting the feature map resolution, a 3×3 convolution smoothing process is applied to each feature map. Finally, the adjusted feature map is combined with the enhanced output features of all levels through the Hadamard product, thereby enhancing the semantic information and detail information of the output features, and obtaining the enhanced multi-scale camouflaged target edge information f out , the specific calculation formula of the enhanced multi-scale camouflaged target edge information is: f out =Hadamard(f′ o1 ,f′ o2 ,f′ o3 ,f′ o4 ,f′ o5 )。 9. A camouflaged target segmentation method based on multi-scale boundary fusion, characterized in that: include: Divide the camouflaged target image dataset into a training dataset; Using any image in the training data set as a feature map, the backbone network extracts features from the feature map to obtain multi-level hierarchical feature information, which contains low-level features containing detail information and high-level features containing semantic information; Combining low-level features and high-level features, and utilizing a parameter-free attention mechanism to enhance the camouflaged target region information in the feature map, and extracting the camouflaged target edge information based on the enhanced camouflaged target region information; Global average pooling is used on all levels of feature information to extract multi-scale edge feature information. This information is combined with the camouflaged target edge information and enhanced through fine-grained detail feature enhancement to obtain enhanced detail features. This enhanced detail feature is then combined with the camouflaged target edge information to obtain the enhanced output features of each level. The Hadamard product is used to combine the enhanced output features of all levels to obtain the enhanced multi-scale camouflaged target edge information; Using the positioning module and multi-level focusing module in PFNet, the enhanced multi-scale camouflaged target edge information is combined with the low-level features, and the combined information is passed sequentially in the multi-level focusing modules. The prediction result of the last level focusing module is used as the final segmentation result.

10. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to claim 9 are implemented.

Images