A method and system for camouflaged target detection based on boundary interaction learning

Abstract

This invention discloses a camouflage target detection method and system based on boundary interaction learning, belonging to the field of camouflage target detection technology. The method includes: S1: Preprocessing the image to be detected and inputting it into an encoder, extracting multi-level encoded features through convolution operations; S2: Based on the multi-level encoded features, extracting high-frequency information through a Laplacian pyramid and fusing it with deep encoded features to generate multi-scale boundary features; S3: Generating boundary interaction features through bidirectional interaction operations, and performing attention weighting and calibration operations on the boundary interaction features to generate global contour enhancement features; S4: Based on the multi-level encoded features and decoded features, generating local optimization features through semantic guidance and spatial detail extraction operations; S5: Performing fusion operations based on global contour enhancement features and local optimization features, and upsampling and feature concatenation through a decoder to generate a camouflage target semantic segmentation result. This enhances the robustness and accuracy of the detection model.

Description

Technical Field

[0001] This invention relates to the field of camouflage target detection technology, and in particular to a camouflage target detection method and system based on boundary interaction learning. Background Technology

[0002] Camouflage target detection is a key task in computer vision, aiming to detect and track targets with camouflage characteristics from the background, with applications in military and wildlife monitoring. However, camouflage target detection under complex background conditions faces many challenges. First, due to the similarity between camouflage targets and the background, the boundary features of camouflage targets are blurred, making it difficult for traditional methods to accurately distinguish between camouflage targets and the background. Second, in practical applications, existing deep learning-based camouflage target detection methods typically have high computational complexity, making it difficult to run efficiently in resource-constrained scenarios.

[0003] Current methods for camouflaged target detection have demonstrated that boundaries serve as effective auxiliary cues, aiding in the detection of the overall structure of camouflaged targets. Introducing additional boundary information can guide feature learning for camouflaged targets, thereby promoting the generation of global features that highlight their structure. However, these methods rely solely on high-frequency information as unidirectional prior guidance, making them susceptible to noise.

[0004] Furthermore, if the global features of the target are overemphasized without considering local details, it is easy to confuse the detailed features between the target and the background.

[0005] Due to the inherent limitations of camouflaged targets and frequent obstruction by background clutter, camouflaged target detectors face ongoing challenges. Therefore, how to achieve semantic segmentation of such objects in complex real-world scenarios is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] This invention provides a camouflaged target detection method and system based on boundary interaction learning to solve the above-mentioned problems in the prior art.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A camouflaged target detection method based on boundary interaction learning includes:

[0009] S1: The image to be detected is preprocessed and then input into the encoder to extract multi-level coding features through convolution operations;

[0010] S2: Based on multi-level coding features, construct a Laplacian pyramid, and convolve the high-frequency features of each layer of the Laplacian pyramid and the deep features of the encoder respectively, and then concatenate them in the channel dimension to generate multi-scale boundary features.

[0011] S3: Based on multi-scale boundary features and multi-level coding features, boundary interaction features are generated through bidirectional interactive operations. Attention weighting and calibration operations are performed on the boundary interaction features to generate global contour enhancement features.

[0012] S4: Based on multi-level encoding and decoding features, local optimized features are generated through semantic guidance and spatial detail extraction operations;

[0013] S5: Based on the fusion operation of global contour enhancement features and local optimization features, the decoder performs upsampling and feature concatenation to generate the semantic segmentation result of the disguised target.

[0014] Furthermore, step S2 includes:

[0015] S21: Perform Gaussian filtering and downsampling operations on the input image to construct a Gaussian pyramid;

[0016] S22: Based on the Gaussian pyramid, high-frequency features of each layer are extracted through differential operations between adjacent layers to construct the Laplace pyramid;

[0017] S23: Based on the high-frequency features of each layer of the Laplacian pyramid and the deep features of the encoder, convolution operations are performed separately and then the data is concatenated in the channel dimension.

[0018] S24: Based on the splicing results, convolution, normalization and nonlinear transformation operations are performed sequentially to generate multi-scale boundary features.

[0019] Furthermore, step S3 includes:

[0020] S31: Based on the current layer multi-scale boundary features, the previous layer boundary interaction features and the current layer target features, enhanced boundary features are generated through multiplication and residual connection. The enhanced boundary features are then spliced ​​and fused to generate the current layer boundary interaction features.

[0021] S32: Based on the current layer boundary interaction features, multi-level encoding features and the previous layer decoding features, an attention map is generated through reverse operation and boundary extraction operation. The encoding features are multiplied with the attention map and then concatenated to generate attention features.

[0022] S33: Based on attention features, an attention mask is generated through convolution and sigmoid operations. The attention features are multiplied by the attention mask and then added to the encoded features to generate attention-weighted features.

[0023] S34: Based on attention-weighted features, channel and spatial calibration operations are performed through the convolutional block attention module to generate global contour enhancement features.

[0024] Furthermore, the k-th layer image of the Gaussian pyramid is generated as follows: when k is 0, the k-th layer image is the input image; when k is greater than 0, the (k-1)-th layer image is subjected to Gaussian filtering and then downsampled by a factor of 2 to generate the k-th layer image.

[0025] The features of the k-th layer of the Laplacian pyramid are generated as follows: the image of the (k+1)-th layer is upsampled to the size of the k-th layer, and the k-th layer image is subtracted from the upsampled result to generate the high-frequency features of the k-th layer;

[0026] Where k represents the pyramid level index.

[0027] Furthermore, in step S31:

[0028] Multi-scale boundary features of the current layer With the target features of the current layer Multiplication, and then combining the results with the multiplication result. Add them together to generate the first enhanced boundary feature;

[0029] Interactive features of the upper layer boundary and Multiplication, and then combining the results with the multiplication result. Add them together to generate a second enhanced boundary feature;

[0030] The first and second enhanced boundary features are concatenated and then fused to generate the current layer boundary interaction features. ;

[0031] Where i represents the layer index, and the fusion operations include convolution, normalization, and ReLU operations.

[0032] Furthermore, in step S32:

[0033] Decoding features of the previous layer For each element, perform the operation of subtracting the element's value from 1 to generate a reverse attention graph;

[0034] right Perform Laplacian operator convolution to generate a boundary attention map;

[0035] Multi-level encoding features Interacting features with the inverse attention map, boundary attention map, and current layer boundary respectively. Multiplying them generates three weighted features;

[0036] The three weighted features are concatenated and then convolved to generate attention features. .

[0037] Furthermore, step S4 includes:

[0038] S41: Based on the decoding features of the previous layer, semantic guidance information is extracted through convolution operations;

[0039] S42: Based on semantic guidance information and multi-level encoded features, spatial detail features are extracted through convolution and channel attention operations;

[0040] S43: Integrating spatial detail features with learnable parameters Multiplication: The result of multiplication is added to the multi-level encoded features to generate locally optimized features;

[0041] Among them, learnable parameters Obtained through backpropagation training.

[0042] Furthermore, step S5 includes:

[0043] S51: Integrating global contour enhancement features with learnable parameters Multiplication combines the locally optimized features with the learnable parameters. Multiplication adds the results of two multiplications to generate aggregated features;

[0044] S52: Perform convolution, normalization, and ReLU operations on the aggregated features to generate the current layer's decoded features;

[0045] S53: Upsample the current layer's decoded features by a factor of two, and concatenate the upsampled results with the corresponding coding layer features;

[0046] S54: Repeat S52 and S53 until the input resolution is restored, perform convolution operation on the final result, and generate the camouflaged target semantic segmentation result;

[0047] Among them, learnable parameters and Obtained through backpropagation training.

[0048] Furthermore, a camouflaged target detection system based on boundary interaction learning applied to the method includes:

[0049] The encoder module is used to perform multi-layer convolution operations on the preprocessed image to be detected to generate multi-layer encoded features at different resolutions.

[0050] The multi-scale boundary feature extraction module is used to extract high-frequency features through the Laplacian pyramid, and then splice and fuse the high-frequency features with the deep features of the encoder to generate multi-scale boundary features.

[0051] The boundary interaction learning module is used to generate boundary interaction features by multiplying boundary features with target features and performing residual operations. It also generates global contour enhancement features through attention weighting and convolutional block attention calibration.

[0052] The local feature refinement module is used to generate semantic guidance information based on decoded features. It extracts spatial details through semantic guidance and encoded features and connects them with the residual of encoded features to generate local optimized features.

[0053] The decoder module is used to perform weighted fusion of global contour enhancement features and local optimization features, and restore spatial resolution through upsampling and feature concatenation to generate semantic segmentation results for camouflaged targets.

[0054] Compared with the prior art, the present invention has the following advantages:

[0055] This invention discloses a camouflage target detection method and system based on boundary interaction learning. The method utilizes a multi-scale boundary extraction module to fuse high-frequency and high-level semantic information, capturing target boundary features at different scales to generate more refined boundary information and avoid information loss at a single scale. Through a boundary interaction learning module, a bidirectional interaction mechanism deeply fuses boundary information with target features, enhancing the model's understanding of target shape and location and gradually generating more refined global features. A local feature refinement module further optimizes the local feature representation of the target. This overall strategy considers both the global outline and local details of the camouflage target, generating more complete and discriminative camouflage target features, thereby improving the detection accuracy of camouflage targets. Attached Figure Description

[0056] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0057] Figure 1 This is a flowchart of a camouflage target detection method based on boundary interaction learning in an embodiment of the present invention;

[0058] Figure 2 This is a schematic diagram of the multi-scale boundary extraction module provided in an embodiment of the present invention;

[0059] Figure 3 This is a schematic diagram of the boundary interaction learning module in an embodiment of the present invention;

[0060] Figure 4 This is a schematic diagram of the local feature refinement module in an embodiment of the present invention. Detailed Implementation

[0061] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0062] The embodiments of the present invention provide, as follows Figure 1As shown, a camouflage target detection method based on boundary interaction learning includes:

[0063] S1: The image to be detected is preprocessed and then input into the encoder to extract multi-level coding features through convolution operations;

[0064] S2: Based on multi-level coding features, construct a Laplacian pyramid, and convolve the high-frequency features of each layer of the Laplacian pyramid and the deep features of the encoder respectively, and then concatenate them in the channel dimension to generate multi-scale boundary features.

[0065] S3: Based on multi-scale boundary features and multi-level coding features, boundary interaction features are generated through bidirectional interactive operations. Attention weighting and calibration operations are performed on the boundary interaction features to generate global contour enhancement features.

[0066] S4: Based on multi-level encoding and decoding features, local optimized features are generated through semantic guidance and spatial detail extraction operations;

[0067] S5: Based on the fusion operation of global contour enhancement features and local optimization features, the decoder performs upsampling and feature concatenation to generate the semantic segmentation result of the disguised target.

[0068] The following is a detailed description with reference to specific embodiments.

[0069] Example 1:

[0070] This embodiment provides a camouflaged target detection method based on boundary interaction learning, including the following steps:

[0071] S1: The image to be detected is preprocessed and then input into the encoder to extract multi-level coding features through convolution operations;

[0072] The image to be detected undergoes preprocessing, including data augmentation operations such as random rotation, scaling, and cropping, to uniformly adjust the image to a fixed size. The preprocessed image is then input into an encoder (such as a ResNet or PVT backbone network), where multi-level convolutional operations extract multi-scale, multi-level encoded features at different scales. ,in This represents the encoded feature of the i-th layer. As the feature resolution decreases, the semantic information gradually increases.

[0073] S2: Based on multi-level coding features, construct a Laplacian pyramid, and convolve the high-frequency features of each layer of the Laplacian pyramid and the deep features of the encoder respectively, and then concatenate them in the channel dimension to generate multi-scale boundary features.

[0074] like Figure 2As shown, based on the multi-level coding features extracted in step S1, high-frequency boundary information is extracted through the Laplacian pyramid and fused with the deep coding features to generate multi-scale boundary features, specifically including:

[0075] S21: Construct a Gaussian pyramid from the input image, and generate a multi-scale image representation through Gaussian filtering and downsampling operations. (Image of the k-th layer of the Gaussian pyramid) The generation method is as follows:

[0076] ;

[0077] in, For the input image, This indicates a Gaussian filtering operation. This indicates a 2x downsampling operation.

[0078] S22: Construct a Laplacian pyramid based on a Gaussian pyramid, and extract high-frequency features of each layer through difference operations between adjacent layers. Features of the k-th layer of the Laplacian pyramid. Generate in the following way:

[0079] ;

[0080] in, This indicates an upsampling operation, which upsamples the (k+1)th layer image to the kth layer size. The difference between the two is used to obtain the high-frequency features of the kth layer, which contain boundary information at that scale.

[0081] S23: Based on the high-frequency characteristics of each layer of the Laplace pyramid Deep coding features extracted by the encoder , respectively and Perform a 1×1 convolution operation to adjust the number of channels, and then perform a concatenation operation along the channel dimension.

[0082] S24: Based on the concatenation result, perform convolution, batch normalization, and ReLU nonlinear transformation operations sequentially to generate multi-scale boundary features. The calculation formula is:

[0083] ;

[0084] in, This represents a 1×1 convolution operation. This indicates concatenation along the channel dimension. This represents a series of operations involving convolution, batch normalization, and the ReLU activation function.

[0085] S3: Based on multi-scale boundary features and multi-level coding features, boundary interaction features are generated through bidirectional interactive operations. Attention weighting and calibration operations are performed on the boundary interaction features to generate global contour enhancement features.

[0086] like Figure 3 As shown, based on multi-scale boundary features and multi-level coding features, boundary interaction features are generated through bidirectional interaction operations. Attention weighting and calibration operations are then performed on the boundary interaction features to generate global contour enhancement features, specifically including:

[0087] S31: Based on the multi-scale boundary features of the current layer Upper-layer boundary interaction features and current layer target features Enhanced boundary features are generated through multiplication and residual connections. These enhanced boundary features are then concatenated and fused to generate the current layer's boundary interaction features. .

[0088] Specifically, the multi-scale boundary features of the current layer With the target features of the current layer Perform element-wise multiplication, then perform a fusion operation (including convolution, normalization, and ReLU) on the multiplication results before combining them with... Add them together to generate the first enhanced boundary feature:

[0089] ;

[0090] At the same time, the boundary interaction features of the upper layer will be... and Perform element-wise multiplication, then combine the results and add them together. Add them together to generate a second enhanced boundary feature:

[0091] ;

[0092] The first and second enhanced boundary features are concatenated along the channel dimension and then fused to generate the current layer boundary interaction features:

[0093] ;

[0094] in, This indicates element-wise multiplication. Indicates fusion operation, This indicates a feature splicing operation.

[0095] S32: Based on the current layer boundary interaction features Multi-level coding features and the decoding features of the previous layer An attention map is generated through inverse operations and boundary extraction operations. The encoded features are then multiplied by the attention map and concatenated to generate attention features. .

[0096] Specifically, for the decoding features of the previous layer For each element, subtract 1 from its value to generate a reverse attention graph:

[0097] ;

[0098] right Perform the Laplace operator Convolution operation extracts boundary information and generates a boundary attention map:

[0099] ;

[0100] Multi-level encoding features Interacting features with the inverse attention map, boundary attention map, and current layer boundary respectively. Element-wise multiplication is performed to generate three weighted features. These three weighted features are then concatenated and convolved to generate attention features.

[0101] ;

[0102] Where [] denotes feature concatenation operation, This represents the Laplace operator.

[0103] S33: Based on attention features Attention masks are generated through convolution and sigmoid operations. The attention features are multiplied element-wise with the attention mask and then combined with the encoded features. Add them together to generate attention-weighted features. The purpose of attention masks is to guide the model to focus its attention on important regions while suppressing background noise and redundant information.

[0104] ;

[0105] ;

[0106] in, This represents the sigmoid activation function.

[0107] S34: Attention-Weighted Features Global contour enhancement features are generated by performing channel and spatial dimension calibration operations through a convolutional block attention module (CBAM). :

[0108] ;

[0109] The CBAM module sequentially applies channel attention and spatial attention weights to the features, enhancing the response of important feature channels and spatial locations.

[0110] S4: Based on multi-level encoding and decoding features, local optimized features are generated through semantic guidance and spatial detail extraction operations;

[0111] like Figure 4 As shown, based on multi-level coding features and decoding features By using semantic guidance and spatial detail extraction operations, locally optimized features are generated. Specifically, it includes:

[0112] S41: Based on the decoding features of the previous layer Semantic guidance information is extracted through convolution operations. This semantic information contains prior knowledge of the target region learned by the upper-layer decoder.

[0113] S42: Based on semantic guidance information and multi-level coding features Spatial detail features are extracted through convolution and channel attention operations. A Local Feature Refinement (LRM) module, guided by the decoded features, extracts key spatial details from the encoded features while filtering out noisy channels.

[0114] S43: Integrating spatial detail features with learnable parameters Multiplication, combining the result with multi-level encoded features. Perform residual connection summation to generate local optimization features:

[0115] ;

[0116] Among them, learnable parameters The weights are automatically learned through backpropagation training and used to adjust the weights of local refinement features.

[0117] S5: Based on the fusion operation of global contour enhancement features and local optimization features, the decoder performs upsampling and feature concatenation to generate the semantic segmentation result of the camouflaged target;

[0118] Based on global contour enhancement features and local optimization features The fusion operation involves upsampling and feature concatenation via a decoder to generate a semantic segmentation result for the disguised target. Specifically, this includes:

[0119] S51: Enhance global contour features With learnable parameters Multiplication will optimize the local features. With learnable parameters Multiplication adds the results of two multiplications to generate aggregated features. This operation achieves adaptive weighted fusion of global contour information and local detail information.

[0120] S52: Perform fusion operations on the aggregated features, including 3×3 convolution, batch normalization, and ReLU nonlinear transformation, to generate the decoding features of the current layer. :

[0121] ;

[0122] Among them, learnable parameters and It is learned automatically through backpropagation training.

[0123] S53: Decode features for the current layer A 2x upsampling operation is performed, and the upsampling result is concatenated with the corresponding coding layer features along the channel dimension to achieve cross-layer feature fusion.

[0124] S54: Repeat steps S52 and S53, decoding layer by layer until the feature map is restored to the original resolution of the input image. Perform a 1×1 convolution operation on the final decoding result to map the multi-channel features into a single-channel segmentation prediction map, generating a semantic segmentation result for the camouflaged target. In the segmentation result, each pixel value represents the probability that the location belongs to the camouflaged target.

[0125] During model training, a weighted combination of the binary cross-entropy loss function and the IoU loss function is used as the total loss function, and the network parameters are optimized through backpropagation. The training data comes from a publicly available camouflage target detection dataset, which includes camouflage targets and their pixel-level annotations in various scenarios.

[0126] Example 2

[0127] This embodiment provides a camouflaged target detection system based on boundary interaction learning that applies the above method. The system includes the following modules:

[0128] The encoder module performs multi-layer convolution operations on the preprocessed image to be detected, generating multi-level encoded features at different resolutions. This module uses a backbone network (such as ResNet-50 or PVT-v2) as a feature extractor, and outputs encoded features at four different scales through layer-by-layer convolution and pooling operations. As the network depth increases, the spatial resolution of features gradually decreases, while the level of semantic abstraction gradually increases. The multi-level features extracted by the encoder module contain both shallow spatial detail information and deep high-level semantic information.

[0129] The multi-scale boundary feature extraction module extracts high-frequency boundary features using a Laplacian pyramid, concatenating and fusing these high-frequency features with deep encoder features to generate multi-scale boundary features. This module first constructs a Gaussian pyramid on the input image, generating a multi-scale image representation through Gaussian filtering and downsampling operations. Then, a Laplacian pyramid is constructed based on the difference operations between adjacent layers of the Gaussian pyramid to extract high-frequency boundary features at different scales. .

[0130] High-frequency characteristics of each layer of the Laplace pyramid Deep features of the encoder After performing 1×1 convolution operations on each channel, the data is concatenated along the channel dimension. Then, multi-scale boundary features are generated by sequentially applying convolution, batch normalization, and ReLU activation. This module effectively integrates high-frequency boundary information and deep semantic information, providing a boundary representation with global contour awareness capabilities for subsequent boundary interaction learning.

[0131] The boundary interaction learning module generates boundary interaction features through multiplication of boundary features and target features and residual operations. It then generates global contour enhancement features through attention weighting and convolutional block attention calibration. This module implements a bidirectional interaction learning mechanism between boundary features and target features. First, it processes the multi-scale boundary features of the current layer... Interaction features with the upper layer boundary Each with the target features of the current layer Element-wise multiplication and fusion are performed, and enhanced boundary features are generated through residual connections. These features are then concatenated and fused to generate the current layer's boundary interaction features. .

[0132] Secondly, the features decoded in the previous layer The process involves decomposition to generate inverse attention maps and boundary attention maps. The encoded features are then processed. The attention features are generated by element-wise multiplication with the two attention maps and the boundary interaction features, concatenating them, and then performing convolution. .

[0133] Furthermore, attention masks are generated based on attention features through convolution and sigmoid operations. The attention-weighted features are generated by multiplying the attention features with the mask and then concatenating them with the residual of the encoded features. .

[0134] Finally, the attention-weighted features are input into the convolutional block attention module (CBAM) for channel and spatial dimension recalibration, outputting global contour enhancement features. This module optimizes target features through boundary guidance, while simultaneously enhancing boundary representation using target features, gradually generating global features with fine-grained contour information.

[0135] like Figure 3 As shown, the local feature refinement module is used to generate semantic guidance information based on the decoded features. It extracts spatial details through semantic guidance and encoded features, and connects them with the encoded feature residuals to generate locally optimized features. This module builds upon the previous layer of decoded features. As semantic guidance, semantic guidance information containing prior knowledge of the target region is extracted through convolutional operations. Based on this semantic guidance information and the encoded features of the current layer... By using the convolution and channel attention operations of the Local Feature Refinement Module (LRM), key spatial detail features are extracted while background noise channels in the encoded features are filtered out.

[0136] Extracted spatial detail features and learnable parameters Multiply, then combine with the original encoded features Perform residual connections to generate local optimization features. This module, based on global contour guidance, further refines the local features of the camouflaged target, enhancing the model's ability to express detailed information.

[0137] The decoder module performs weighted fusion of global contour enhancement features and local optimization features, restoring spatial resolution through upsampling and feature concatenation to generate semantic segmentation results for camouflaged targets. This module first performs weighted fusion of global contour enhancement features... With learnable parameters Multiplication will optimize the local features. With learnable parameters The two features are multiplied and added together to generate aggregated features. A 3×3 convolution, batch normalization, and ReLU operation are then performed on the aggregated features to generate the current layer's decoding features. .

[0138] The current layer's decoded features are upsampled by a factor of 2. The upsampled result is then concatenated with the corresponding coding layer's features along the channel dimension to achieve cross-layer feature fusion. This fusion and upsampling operation is repeated layer by layer upwards until the feature map is restored to the original resolution of the input image.

[0139] A 1×1 convolution operation is performed on the final decoded features to map the multi-channel features into a single-channel segmentation prediction map, outputting the semantic segmentation result of the disguised target. Each pixel value in the output represents the probability that the location belongs to the disguised target. Typically, a threshold of 0.5 is set for binarization to obtain the final disguised target segmentation mask.

[0140] This system extracts multi-level features through an encoder, obtains global contour information through a multi-scale boundary feature extraction module, achieves bidirectional enhancement of the boundary and the target through a boundary interaction learning module, optimizes detailed representation by combining a local feature refinement module, and finally outputs accurate segmentation results by fusing global and local features through a decoder module. This effectively solves the problem of distinguishing camouflaged targets from complex backgrounds and improves the robustness and accuracy of the detection model.

[0141] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from the spirit and scope of this invention.

Claims

1. A camouflage target detection method based on boundary interaction learning, characterized in that, include: S1: The image to be detected is preprocessed and then input into the encoder to extract multi-level coding features through convolution operations; S2: Based on multi-level coding features, construct a Laplacian pyramid, and convolve the high-frequency features of each layer of the Laplacian pyramid and the deep features of the encoder respectively, and then concatenate them in the channel dimension to generate multi-scale boundary features. S3: Based on multi-scale boundary features and multi-level encoded features, boundary interaction features are generated through bidirectional interactive operations. Attention weighting and calibration operations are then performed on these boundary interaction features to generate global contour enhancement features; including: S31: Based on the current layer multi-scale boundary features, the previous layer boundary interaction features and the current layer target features, enhanced boundary features are generated through multiplication and residual connection. The enhanced boundary features are then spliced ​​and fused to generate the current layer boundary interaction features. S32: Based on the current layer boundary interaction features, multi-level encoding features and the previous layer decoding features, an attention map is generated through reverse operation and boundary extraction operation. The encoding features are multiplied with the attention map and then concatenated to generate attention features. S33: Based on attention features, an attention mask is generated through convolution and sigmoid operations. The attention features are multiplied by the attention mask and then added to the encoded features to generate attention-weighted features. S34: Based on attention-weighted features, channel and spatial calibration operations are performed through a convolutional block attention module to generate global contour enhancement features. S4: Based on multi-level encoding and decoding features, local optimized features are generated through semantic guidance and spatial detail extraction operations; S5: Based on the fusion operation of global contour enhancement features and local optimization features, the decoder performs upsampling and feature concatenation to generate the semantic segmentation result of the disguised target.

2. The camouflage target detection method based on boundary interaction learning according to claim 1, characterized in that, Step S2 includes: S21: Perform Gaussian filtering and downsampling operations on the input image to construct a Gaussian pyramid; S22: Based on the Gaussian pyramid, high-frequency features of each layer are extracted through differential operations between adjacent layers to construct the Laplace pyramid; S23: Based on the high-frequency features of each layer of the Laplacian pyramid and the deep features of the encoder, convolution operations are performed separately and then the data is concatenated in the channel dimension. S24: Based on the splicing results, convolution, normalization and nonlinear transformation operations are performed sequentially to generate multi-scale boundary features.

3. The camouflaged target detection method based on boundary interaction learning according to claim 2, characterized in that, The image of the k-th layer of the Gaussian pyramid is generated as follows: when k is 0, the image of the k-th layer is the input image; when k is greater than 0, the image of the (k-1)-th layer is subjected to Gaussian filtering and then downsampled by a factor of 2 to generate the image of the k-th layer. The features of the k-th layer of the Laplacian pyramid are generated as follows: the image of the (k+1)-th layer is upsampled to the size of the k-th layer, and the k-th layer image is subtracted from the upsampled result to generate the high-frequency features of the k-th layer; Where k represents the pyramid level index.

4. The camouflaged target detection method based on boundary interaction learning according to claim 1, characterized in that, In step S31: Multi-scale boundary features of the current layer With the target features of the current layer Multiplication, and then combining the results with the multiplication result. Add them together to generate the first enhanced boundary feature; Interactive features of the upper layer boundary and Multiplication, and then combining the results with the multiplication result. Add them together to generate a second enhanced boundary feature; The first and second enhanced boundary features are concatenated and then fused to generate the current layer boundary interaction features. ; Where i represents the layer index, and the fusion operations include convolution, normalization, and ReLU operations.

5. The camouflaged target detection method based on boundary interaction learning according to claim 1, characterized in that, In step S32: Decoding features of the previous layer For each element, perform the operation of subtracting the element's value from 1 to generate a reverse attention graph; right Perform Laplacian operator convolution to generate a boundary attention map; Multi-level encoding features Interacting features with the inverse attention map, boundary attention map, and current layer boundary respectively. Multiplying them generates three weighted features; The three weighted features are concatenated and then convolved to generate attention features. .

6. The camouflaged target detection method based on boundary interaction learning according to claim 1, characterized in that, Step S4 includes: S41: Based on the decoding features of the previous layer, semantic guidance information is extracted through convolution operations; S42: Based on semantic guidance information and multi-level encoded features, spatial detail features are extracted through convolution and channel attention operations; S43: Integrating spatial detail features with learnable parameters Multiplication: The result of multiplication is added to the multi-level encoded features to generate locally optimized features; Among them, learnable parameters Obtained through backpropagation training.

7. The camouflage target detection method based on boundary interaction learning according to claim 1, characterized in that, The S5 steps include: S51: Integrating global contour enhancement features with learnable parameters Multiplication combines the locally optimized features with the learnable parameters. Multiplication adds the results of two multiplications to generate aggregated features; S52: Perform convolution, normalization, and ReLU operations on the aggregated features to generate the current layer's decoded features; S53: Upsample the current layer's decoded features by a factor of two, and concatenate the upsampled results with the corresponding coding layer features; S54: Repeat S52 and S53 until the input resolution is restored, perform convolution operation on the final result, and generate the camouflaged target semantic segmentation result; Among them, learnable parameters and Obtained through backpropagation training.

8. A camouflaged target detection system based on boundary interaction learning applied to the method of claim 1, characterized in that, include: The encoder module is used to perform multi-layer convolution operations on the preprocessed image to be detected to generate multi-layer encoded features at different resolutions. The multi-scale boundary feature extraction module is used to extract high-frequency features through the Laplacian pyramid, and then splice and fuse the high-frequency features with the deep features of the encoder to generate multi-scale boundary features. The boundary interaction learning module is used to generate boundary interaction features by multiplying boundary features with target features and performing residual operations. It also generates global contour enhancement features through attention weighting and convolutional block attention calibration. The local feature refinement module is used to generate semantic guidance information based on decoded features. It extracts spatial details through semantic guidance and encoded features and connects them with the residual of encoded features to generate local optimized features. The decoder module is used to perform weighted fusion of global contour enhancement features and local optimization features, and restore spatial resolution through upsampling and feature concatenation to generate semantic segmentation results for camouflaged targets.

Images