Camouflage target detection method based on feature selection attention and frequency domain edge guidance

Abstract

The invention discloses a camouflage target detection method based on feature selection attention and frequency domain edge guidance. According to the method, four-level features of a camouflage target image are extracted through a backbone network SMT and are respectively screened; the high-level features are input into a semantic information supplement module, and after semantic features are enhanced, the high-level features and the trunk features are sent into a spatial feature enhancement module together. And inputting the obtained fine-grained features into an edge feature sensing module, and finally fusing multi-scale features through a multi-scale jump connection technology to generate a mask pattern with higher discrimination. The method has the advantages that the network parameter quantity is reduced and key information is reserved through a feature selection mechanism; a spatial feature enhancement module is used for enhancing multi-scale feature representation and remote dependence modeling; the dilution of the semantic context is relieved by means of a semantic supplement module so as to improve the positioning precision; and an edge feature enhancement module is adopted to enhance edge semantic perception and improve boundary integrity. According to the method, the camouflage target detection performance is remarkably improved with relatively low calculation cost.

Description

TECHNICAL FIELD

[0001] The present application relates to a camouflage target detection method based on feature selection attention and frequency domain edge guidance, which is a target detection method for camouflage scenes and belongs to the cross field of engineering application and information science. BACKGROUND

[0002] Camouflage target detection refers to the task of detecting targets embedded in the environment through self-camouflage or environmental occlusion. The main goal of camouflage target detection is to distinguish the subtle differences between complex background environments and target objects with camouflage features, and to completely identify the target objects from the background. The core difficulty lies in capturing the extremely low visual distinguishable differences between the foreground and the background. Existing methods can be divided into three categories: traditional algorithms based on handcrafted features, deep learning methods based on convolutional neural networks (CNN), and deep learning methods based on Transformers.

[0003] Traditional camouflage target detection methods are based on handcrafted features, including texture feature extraction, detection of convex intensity using hand-designed operators, edge and color extraction, and then selecting a classifier to generate a prediction mask. Although this type of solution has interpretability, it requires a large amount of domain prior knowledge and the sub-feature description capability is limited, making it difficult to generalize to complex and variable natural scenes, which restricts large-scale applications.

[0004] Camouflage target detection methods based on CNN use a pre-trained CNN backbone to extract multi-level coarse features of images, and then use a carefully designed convolution module for refinement and fusion, significantly improving detection accuracy. However, due to the limitations of local receptive fields, CNNs have difficulty establishing long-distance dependencies, and are prone to missed detection and edge fragmentation for camouflage targets with significant scale differences or sparse distribution.

[0005] Camouflage target detection methods based on Transformers differ from CNN methods in that they use a pre-trained Transformer backbone to extract multi-level coarse features of images. Compared with CNN-based models, Transformers have outstanding performance in long-range context modeling due to their self-attention mechanism, and have gradually become the mainstream backbone for camouflage target detection. However, the parameter quantity and computational complexity are huge, which puts a heavy burden on end-side deployment.

[0006] In existing research, these methods have problems such as insufficient utilization of multi-scale information and structured information, weak ability to establish long-range context dependency, and insufficient differentiation of target edges and background differences. For example, in actual situations, there may be multiple camouflage targets of different sizes in a scene, and the target edges are extremely similar to the background. Existing methods may miss individual targets and have fragmented edge segmentation when detecting. In addition, the mainstream U-shaped architecture has semantic attenuation in the process of transferring high-level semantics to low-level space, causing target positioning drift. At the same time, excessive parameter quantity and computational overhead further hinder the deployment and real-time application of the algorithm in resource-constrained environments. SUMMARY

[0007] The purpose of the invention is to solve the problems existing in the current camouflage target detection method: First, due to the limited receptive field, existing models cannot well explore the dependency relationship between multi-scale long-distance context features, and usually cannot obtain global features, leading to problems such as over-segmentation or local omission in scenes with targets of large size difference. In addition, semantic and global context information may be gradually diluted in the top-down feature transmission path, exacerbating target positioning bias. Second, distinguishing the blurred boundary between camouflage targets and background is the main difficulty of the task. Existing methods have insufficient utilization of high-frequency spatial information such as boundaries and textures, leading to decreased detection performance in scenes with complex blurred boundaries and inability to identify the structure and details of the target edge completely. Third, existing high-performance models usually have large parameters and high computational complexity, which hinders further practical application. The purpose of the invention is to provide a camouflage target detection method based on feature selection attention and frequency domain edge guidance to solve these problems in the current camouflage target detection method.

[0008] Technical solution: To achieve the above object, the application provides a camouflage target detection method based on feature selection attention and frequency domain edge guidance. A new selective multi-scale spatial attention enhancement method is proposed, which uses a feature selection mechanism, a dimension decoupling modeling method and a layered skip-connection receptive field module to fully extract fine-grained multi-scale spatial features and long-distance context structure information in the image, effectively establish long-distance channel dependency, and generate more discriminative feature representations. In order to reduce the semantic loss in the U-shaped network, high-level semantic information is extracted and input into each layer of the spatial feature enhancement module to supplement the semantic context information in the feature stream to accurately locate the camouflage object. At the same time, in order to obtain more accurate boundary information of the camouflage target, the discrete cosine transform is introduced to extract the frequency information of the original image, combined with the edge clues between the mask foreground and background, and the three-way branch dynamic fusion is fully learned to the boundary features. Finally, a multi-scale skip-connection decoding module is designed to aggregate target features of different semantic scales to enhance the representation ability of multi-scale context information, and a more discriminative mask prediction map is generated through real value supervision. The specific technical scheme includes the following steps:

[0009] Step one: feature selection extraction.

[0010] (1) Coarse feature extraction: use the pre-trained model SMT to encode the camouflage image. Through the combination of multi-layer CNN and Transformer architecture, rich context feature information is extracted from the image to generate four different levels of image features to construct multi-level feature representation.

[0011] (2) Feature sorting selection: For the four levels of image features, each level of feature is subjected to average pooling operation in the spatial dimension to obtain the weight information of each group of features in the channel dimension. Then, the weight information is input into the convolution layer for information interaction in the channel dimension, and the original feature is recalibrated with new weight information. The calibrated feature is sorted in the channel dimension according to the weight value, and the top 64 features with the largest weight value are selected for subsequent calculation.

[0012] Step two: semantic information supplement.

[0013] (1) Semantic feature alignment: normalize the size of the top layer feature and the next high layer feature in height and width dimensions, and perform concatenation in channel dimension; then input the fusion result into the convolution layer, perform batch normalization processing, and use the ReLU activation function to generate coarse-grained semantic features.

[0014] (2) Local detail extraction: local detail information of the coarse semantic feature is extracted by deep separable convolution, and after feature calibration using global attention map, deep separable convolution is used again to learn and fuse features to suppress redundant noise, and finally a residual connection is established with the original coarse semantic feature to obtain fine semantic representation of the camouflage target.

[0015] (3) Global semantic aggregation: global average pooling down-sampling is performed on the coarse semantic feature, and one-dimensional convolution operation with high and wide dimension cascade is used to aggregate pixel features to construct a global receptive field; the processing result is input into the Sigmoid function to generate a global attention map, and finally restored to the original feature size by bilinear interpolation. In the subsequent feature calculation stage, the semantic information is injected into each level of the U-shaped network for multi-scale feature fusion.

[0016] Step three: spatial feature enhancement.

[0017] (1) Coarse spatial feature construction: the size of the adjacent two layers of features and semantic features is normalized respectively, and spliced in the channel dimension; then convolution operation, batch normalization and ReLU activation are performed in turn to obtain coarse-grained spatial features.

[0018] (2) Spatial decoupling attention: global spatial information aggregation and local structure information extraction are performed on the coarse spatial features in the height and width dimensions respectively, keeping the position information in the corresponding dimensions and channel dimensions unchanged; the global spatial information and the local structure information are spliced along the channel dimension, and the feature is recalibrated through the excitation operation to capture the dependency between features; the processing result is input into the Sigmoid function to generate a spatial attention map, and multiplied element by element with the original coarse spatial features to obtain fine spatial features.

[0019] (3) Multi-scale hierarchical perception: the spatial features are evenly divided into s sub-layers along the channel dimension, and input into the residual pooling convolution module composed of convolution layers and multiple scale pooling layers to capture multi-scale context information, learn spatial feature expression at different levels, and finally generate spatial enhancement features with local fine-grained features and global context information.

[0020] Step four: edge feature perception.

[0021] (1) Frequency domain information extraction: discrete cosine transform is performed on the original camouflage image to extract frequency domain representation, and high frequency components are cut off, then deep separable convolution and bilinear interpolation are used to scale the feature map, suppress background noise, and generate a frequency attention map through Sigmoid function after channel dimension reduction.

[0022] (2)Edge feature enhancement: the coarse mask is input into the Sigmoid function after size normalization to obtain the foreground attention map, and the background attention map is obtained by logical negation; the spatial enhancement features are respectively multiplied with the frequency attention map, the foreground attention map and the background attention map, the three results are spliced along the channel dimension, and then sequentially subjected to convolution, batch normalization and ReLU activation, and then multiplied with the newly generated Sigmoid attention map, and finally connected with the original spatial enhancement features to output the edge feature.

[0023] (3)Edge mask supervision: the edge feature is reduced to a single-channel edge map through convolution operation, and the real edge is used as a supervision signal to train the camouflage target edge mask.

[0024] Step five: skip connection decoding.

[0025] (1)Hierarchical decoding: the top layer spatial enhancement feature is input into a one-level decoder composed of a convolution layer, a batch normalization layer and a ReLU activation layer, to obtain a top layer decoding feature, and then the decoding feature is input into a two-level decoder together with the next layer edge feature to obtain a two-level decoding feature. The two-level decoding features are spliced along the channel dimension, and the same process is iterated in the next level, and finally four layers of decoding features are output.

[0026] (2)Residual prediction fusion: the channel dimension of each layer of decoding features is reduced to generate a corresponding single-channel mask prediction map, the current layer prediction map is connected with the next layer prediction map, and the real mask is used for supervision at each layer. The prediction maps of all levels are stacked layer by layer to obtain the final mask prediction map.

[0027] Beneficial effects: the present application proposes a camouflage target detection method based on feature selection attention and frequency domain edge guidance. The method can not only establish long-distance dependence on features, but also mine the subtle differences between target edges and backgrounds in the frequency domain, solve the problem of modeling the dependence relationship between long-distance context features of camouflage targets with large scale differences, the problem of semantic feature loss, and the problem of insufficient use of high-frequency spatial information such as boundaries and textures, which leads to performance decline in scenes with complex fuzzy boundaries, resulting in blurred edge structure of the segmentation result. In addition, the model has small number of parameters and low computational complexity, which is beneficial to practical application deployment. BRIEF DESCRIPTION OF DRAWINGS

[0028] Figure 1 The overall flowchart of the camouflage target detection method based on feature selection attention and frequency domain edge guidance.

[0029] Figure 2 is the specific architecture of the model. DETAILED DESCRIPTION

[0030] The present invention will be further described below with reference to the accompanying drawings.

[0031] This invention is a camouflaged target detection method based on feature selection attention and frequency domain edge guidance, the overall flowchart of which is attached. Figure 1 As shown. First, the camouflaged target is input into the pre-trained SMT backbone to extract four levels of multi-scale coarse features, and the subset of channels with the highest relevance to the target is selected layer by layer. Second, the two high-level semantic features are aligned in the spatial dimension and concatenated along the channels. Through parallel modeling of local details and global dependencies, refined high-level semantic features are generated. Subsequently, the semantic features are combined with the four-level original features along the height and width directions to perform global position aggregation and local structure aggregation in the spatial dimension. The channel weights are adaptively recalibrated through activation operations. The sigmoid function generates an attention map and multiplies it element-wise with the input features. Then, it is fed into a residual multi-scale pooling convolution module to output four sets of spatial enhancement features. After dimensionality reduction of the top-level features, a coarse mask prediction map is obtained under the supervision of the real mask. The original image is subjected to discrete cosine transform to decouple and truncate high-frequency components. After noise suppression, a frequency domain attention map is obtained through sigmoid. The coarse mask is sized to generate a foreground attention map, and its inversion yields a background attention map. The three attention maps are multiplied element-wise with the corresponding three-level enhancement features, and three layers of edge enhancement features are formed under edge supervision. Finally, the third-level edge enhancement features and the highest-level enhancement features are decoded through top-down skip connections, fused step by step, and supervised by the mask to output a camouflaged target mask prediction map with clear edges and complete structure. A camouflaged target detection method based on feature selection attention and frequency domain edge guidance mainly consists of five parts: 1) feature selection extraction; 2) semantic information supplementation; 3) spatial feature enhancement; 4) edge feature perception; and 5) skip connection decoding. The specific architecture of the model is shown in the appendix. Figure 2 The specific implementation steps are as follows.

[0032] 1. Feature Selection and Extraction Module

[0033] (1) Coarse feature extraction

[0034] Using backbone network SMT to extract spoofed images Extract coarse features at different levels. The size of features at different levels is represented by H. i ×W i ×C i Among them, H i and W i C represents the height and width of the feature map, respectively. i The number of channels representing the feature, where i represents the i-th layer.

[0035] (2) Feature ranking selection

[0036] For the above coarse features For the i-th level features, a 1x1 global average pooling operation is performed on the spatial dimension to obtain the channel weight information a of each group of features i . Then, the weight information is input into the CBR convolution module for information interaction in the channel dimension. The new weight information is multiplied and calibrated with the original features in the channel dimension. The calibrated features are sorted in the channel dimension according to the weight values. The 64 groups of features with the largest weight values participate in subsequent calculations, and the remaining features do not participate in subsequent processing. The corresponding calculation process is shown in formula (1).

[0037]

[0038] where AvgPool(·) is a 1x1 global average pooling operation, and CBR(·) represents sequentially performing a 3x3 convolution operation, batch normalization, and ReLU activation.

[0039] 2. Semantic information supplement module

[0040] The high-level features of the backbone contain more global semantic information, and making full use of these information helps to enhance the accuracy of locating the camouflaged target. However, since the global semantic information is concentrated in the high-level features, indiscriminately integrating features of all levels will introduce more background noise. Moreover, using convolution-based feature extraction modules is not conducive to establishing long-distance dependencies. The introduction of self-attention operations can help the model capture global information and make up for the shortcomings of convolution-based feature extraction modules in modeling long-distance dependencies. However, typical self-attention operations have too high computational complexity, which is not conducive to the practical deployment of the model. At the same time, the U-shaped structure of the model has a problem of semantic dilution in the top-down feature transmission process, and it is necessary to supplement the high-level semantic information. To solve these problems, this paper designs a semantic information supplement module to aggregate local and global semantic information and capture long-distance dependencies between different spatial pixels.

[0041] (1) Semantic feature alignment

[0042] The bilinear interpolation is used to adjust S3 and S4 to the same size after feature selection and extraction. A 1x1 convolution layer is used to reduce the dimension of the features and adjust the number of channels to C to reduce redundant features and speed up the calculation. After splicing the features of the two branches, the CBR convolution module is used to adjust the feature channels to C to obtain the intermediate feature S in . The specific process can be calculated by formula (2).

[0043] S in =CBR(C at (Conv2d(S3), Conv2d(Up(S4))) (2)

[0044] where Cat(·) denotes the channel concatenation operation, Conv2d(·) denotes a normal 1x1 convolution layer, and Up(·) denotes the bilinear interpolation up-sampling operation.

[0045] (2) Local detail extraction

[0046] Intermediate feature S in enters two parallel branches, i.e., the local information extraction branch and the global information extraction branch. The local branch is composed of a simple 3x3 depthwise separable convolution block to effectively extract the local feature S local of the region of interest with small parameter amount and computational cost to alleviate the problem of local detail loss. The local branch can be calculated by formula (3).

[0047] S local = DSConv(S in ) (3)

[0048] where DSConv(·) denotes a 3x3 depthwise separable convolution block.

[0049] (3) Global semantic aggregation

[0050] The global information extraction branch uses a horizontal convolution layer XConv and a vertical convolution layer YConv to aggregate the feature map S in in the W direction and the H direction, and the two stacks can generate a global receptive field. In addition, an average pooling operation is applied to down-sample the feature map before global information extraction, because it is beneficial to improve efficiency and reduce computational cost to operate small size feature maps. After the global information extraction is completed, the Sigmoid function is used to generate the attention map of the global information, and then the original size is restored to match the size of the feature map S local obtained by the local branch. The specific process can be calculated by formula (4).

[0051] S′ in = YConv(XConv(AvgPool(S in ))) (4)

[0052] S global = Up(σ(S′ in ))

[0053] where S′ in is the intermediate feature, YConv(·) and XConv(·) respectively denote a horizontal convolution layer with a convolution kernel size of 1xk and a vertical convolution layer with a convolution kernel size of kx1, AvgPool(·) denotes an average pooling operation with a reduction rate of r, σ(·) denotes the Sigmoid function, and S global denotes the global feature attention map.

[0054] The attention mask obtained by the global branch is element-wise multiplied with the local feature obtained by the local branch to obtain S out to calibrate local information, model long-distance dependency, and reduce redundant information to suppress background noise. Subsequently, S out Another deep separable convolution is learned to fuse features to obtain more robust feature representation, and a residual connection is added to improve the stability of the model. The specific process can be calculated by formula (5).

[0055]

[0056] wherein the symbol represents element-wise multiplication, and F5 is the final output feature of the semantic information module.

[0057] 3. Spatial feature enhancement module

[0058] (1) Coarse spatial feature construction

[0059] The size of the adjacent two layers of features and the semantic features is normalized respectively, and splicing is completed in the channel dimension; then convolution operation, batch normalization and ReLU activation are sequentially performed to obtain coarse-grained spatial features

[0060]

[0061] (2) Spatial decoupling attention

[0062] The input features fi aggregate features along both height (H) and width (W) directions, which can capture long-range dependencies along one spatial direction while preserving precise location information along the other direction. However, this way ignores detailed feature representation and structure information in the feature aggregation process. To solve this problem, the coordinate aggregation operation in each direction is decoupled into two adaptive parallel branches, global aggregation and local aggregation. Take the operation in the H direction as an example, the global aggregation branch uses a 1xW global average pooling to obtain a global coordinate attention map, which captures long-range dependencies of global features along the H direction and embeds global location information along the W direction. The local aggregation branch uses a 2xW adaptive average pooling to obtain a local coordinate attention map, which encodes structure information of the local feature map and preserves detailed feature representation of the local region. The outputs of the two branches are concatenated in the channel dimension to fuse global and local features. Then, an excitation operation is used to adaptively recalibrate the features to capture inter-channel dependencies, and finally a Sigmoid function is used to generate the coordinate attention map. Similarly, similar operations are performed in the W direction. Finally, the attention maps of the two directions are multiplied with the input features fi to more accurately locate the exact position of the target of interest and capture detailed structure information. The specific process can be calculated by formula (7).

[0063]

[0064] where fi represents the input feature, represents the generation of a global information coordinate map using a global average pooling operation along the H direction or the W direction, represents the generation of a local information coordinate map using a local average pooling operation along the H direction or the W direction. FC H / W (·) represents an excitation operation on the coordinate map of the two directions respectively, which includes a 1x1 channel reduction convolution with a reduction rate of r, a ReLU activation function, a 1x1 channel expansion convolution, and a Sigmoid function to generate an attention map m i represents the output feature after spatially decoupled attention enhancement.

[0065] (3) Multi-scale hierarchical perception

[0066] The output feature m iIn the channel dimension, it is divided into s parts to reduce the model parameter quantity and computational complexity. In order to obtain a more fine-grained feature representation, this paper constructs a pooling convolution module with a residual connection style. The pooling convolution module is composed of a pooling layer with different kernel sizes and a 3x3 convolution layer. The pooling layer with different kernel sizes is responsible for obtaining multiple receptive fields of different scales, and then capturing multi-scale context information. The 3x3 convolution layer has learnable parameters, which is responsible for learning features of different spatial scales. The residual connection mechanism makes the module form a larger receptive field in the step-by-step expansion of the skip connection. The module with a smaller scale receptive field can capture more details in the scene, and the module with a larger scale receptive field can integrate more context information in the scene. The step-by-step increase of the receptive field scale promotes the smooth transition of the model from local feature extraction to global feature extraction. Finally, the output features of each level of the pooling convolution module are spliced to obtain a combination of more rich local fine-grained features and larger range global context features. The specific process can be calculated by formula (8).

[0067]

[0068] where Split(·) represents the channel separation operation, s represents the number of layers of the original feature divided by the channel dimension, and k represents the kth layer. AvgPool 2k-1 (·) represents the average pooling operation with a kernel size of 2k-1x2k-1. F i represents the output feature of the multi-scale hierarchical perception operation.

[0069] 4. Edge feature perception module

[0070] (1) Frequency domain information extraction

[0071] For the original input image I, the frequency domain information is extracted by using the discrete cosine transform Then it is decoupled and enhanced. Specifically, from F freq separate the high-frequency segment Since the high-frequency segment contains more noise, a depth separable convolution with learnable parameters is used to compress the spatial information, and then up-sampling is performed to restore the original feature size to filter irrelevant background noise. Finally, it is compressed into a single channel to collect all the high-frequency features and recalibrate the prominent features with the Sigmoid function to obtain the high-frequency attention map H

[0072] (2) Edge feature enhancement

[0073] The high-frequency attention map H is multiplied element-wise with the output feature Fi of the multi-scale hierarchical perception operation to obtain the high-frequency information stream However, since the camouflage image often has complex background information, it introduces a lot of cluttered redundant information in the high-frequency features. In order to better let the model focus on exploring the edge area of the camouflage target, this paper uses the coarse prediction mask P i+1 As the input to calibrate the area of interest of the module, the coarse mask is first resized and adjusted to the feature stream F i Keep consistent, then generate foreground attention map with Sigmoid function, and then take the inverse to get background attention map. The foreground and background attention maps are multiplied with Fi element by element, and then input into the CBR module to learn the corresponding foreground and background features. Finally, the high-frequency information stream, the foreground information stream, and the background information stream are spliced along the channel dimension, and a nonlinear mapping convolution block is input to learn the edge features under the joint participation of the high-frequency information stream, the foreground stream, and the background stream. The boundary feature is modeled as a region of interest to the high-frequency region, foreground, and background, and the fusion feature is obtained Since the high-frequency information may contain more useless background noise, the attention map generated by the Sigmoid function is multiplied with the fusion feature to recalibrate the feature and ensure that the model focuses on the edge area. Finally, a residual connection is further adopted to enhance the feature representation. The specific process can be calculated by formula (9).

[0074]

[0075] wherein, represents the foreground stream feature, represents the background stream feature, M i is the output feature of the edge feature perception operation, is the edge prediction map, which is supervised by the boundary ground truth.

[0076] 5. Skip connection decoding

[0077] (1) Hierarchical decoding

[0078] In the decoding process, compared with using only one-time concatenation operation, residual skip connection can gradually aggregate multiple decoder features of different semantic scales, which not only can alleviate the loss of details caused by excessive up-sampling, but also can alleviate the problems of gradient vanishing and overfitting. Specifically, the top-layer spatial enhancement feature is input into a first-level decoder composed of a cascade of “convolution layer-batch normalization layer-RELU activation layer”, to obtain the top-layer decoding feature, and then the decoding feature is input into a second-level decoder together with the next-higher-layer edge feature to obtain the second-level decoding feature. After the two-level decoding features are spliced along the channel dimension, the same process is continued at the next level, and finally four-level decoding features are output.

[0079] (2) Residual prediction fusion

[0080] A 1x1 convolution layer is used for channel dimension reduction for each layer of decoded features to generate the corresponding single-channel mask prediction map. The current layer prediction map is connected with the next layer prediction map in residual connection, and the prediction map of each layer is supervised by the true value. All level prediction maps are stacked layer by layer to obtain the final mask prediction map. The loss function L used for supervision can be calculated by formula (10).

[0081]

[0082] wherein, represents the camouflage prediction map generated from each level of residual prediction fusion, represents the boundary prediction map generated in each level of edge feature perception module, L BCE (·) represents the binary cross entropy loss, L IoU (·) represents the intersection over union loss, L Dice (·) represents the dice loss, G t represents the real mask, G e represents the boundary ground truth map.

Claims

1. A camouflage target detection method based on feature selection attention and frequency domain edge guidance, characterized in that, The method comprises the following steps: (1) feature selection extraction: extracting four layers of features from the camouflage image through a pre-trained model SMT, and selecting and extracting features with high target relevance in each layer; (2) semantic information supplement: aligning and splicing the two high-level features in size and channel dimension, then modeling local and global semantic information through two parallel branches respectively, and extracting high-level semantic features; (3) spatial feature enhancement: aggregating the high-level semantic features and the four layers of features in the height and width dimensions respectively, then using excitation operation to adaptively recalibrate the features, capture the dependence between feature channels, then generate an attention map through a Sigmoid function and multiply it with the initial feature matrix element by element, and finally input the residual connection style pooling convolution module to obtain four groups of enhanced features, and the top feature generates a coarse mask graph through dimension reduction convolution and mask supervision; (4) edge feature perception: first, use discrete cosine transform to extract frequency domain information from the original image and decouple to separate high-frequency features, generate a high-frequency attention map using a Sigmoid function after noise filtering, generate a foreground attention map after the coarse mask graph is calibrated in size, and generate a background attention map after taking the inverse, multiply the three attention maps with the three layers of features element by element, and obtain three layers of edge enhanced features through edge supervision; (5) skip connection decoding: the three layers of edge enhanced features and the highest layer of enhanced features are connected and decoded from top to bottom, and are assisted by mask supervision to generate a clear camouflage target mask graph.

2. The camouflage target detection method based on feature selection attention and frequency domain edge guidance according to claim 1, characterized in that, The feature selection extraction in step (1) comprises: (1-1) using a pre-trained SMT as the backbone, performing forward inference on the camouflage image through CNN and Transformer hybrid coding, outputting four levels of multi-scale features, and constructing hierarchical context representation; (1-2) performing global average pooling on each level of features in the spatial dimension to obtain the weight information of each group of features in the channel dimension. Then, the weight information completes cross-channel information interaction through convolution operation, recalibrates the original features with new weight information, and sorts the calibrated features in the channel dimension according to the weight values. The top 64 feature subsets with the largest weight values are retained for subsequent calculation.

3. The camouflage target detection method based on feature selection attention and frequency domain edge guidance according to claim 1, characterized in that, The semantic information supplement in step (2) comprises: (2-1) aligning the top layer and the second high layer features after screening in the spatial dimension by bilinear interpolation, and splicing along the channel axis. The result is input into the convolution module, and after batch normalization and ReLU activation, a coarse-grained semantic feature is generated; (2-2) performing global average pooling compression on the coarse semantic feature, and then applying one-dimensional convolution in the height-width direction in parallel to expand the receptive field. The global attention weight map is obtained through Sigmoid activation, and is restored to the input resolution through bilinear interpolation; (2-3) extracting local fine-grained clues on the coarse semantic feature using depth separable convolution, weighting after global attention map, and then completing feature fusion and noise suppression through depth separable convolution again. Finally, the refined camouflage target semantic feature is obtained by stacking the original coarse semantic feature in the form of residual. (2-4) The semantic representation is injected into each level of the U-shaped network in a skip connection manner to realize the synergistic fusion of multi-scale semantics and spatial features.

4. The camouflage target detection method based on feature selection attention and frequency domain edge guidance according to claim 1, characterized in that, The spatial feature enhancement in step (3) includes: (3-1) Spatial interpolation alignment and channel splicing are performed on the encoding features of the adjacent two levels and the corresponding semantic features, followed by convolution operation, batch normalization and ReLU activation in turn to obtain coarse-grained spatial features; (3-2) Global spatial information aggregation and local structure information extraction are performed on the coarse spatial features in the height and width dimensions respectively, while keeping the position information in the corresponding other dimensions and channel dimensions unchanged; the global spatial information and the local structure information are spliced along the channel dimension, and the feature is recalibrated through the excitation operation, and then a spatial attention map is generated through Sigmoid, which is multiplied element by element with the coarse spatial features to output fine-grained spatial features; (3-3) The fine spatial features are evenly divided into s groups along the channel, and input into the residual pooling-convolution module composed of parallel multi-scale pooling and convolution to capture local-global multi-scale context and output spatial enhancement features.

5. The camouflage target detection method based on feature selection attention and frequency domain edge guidance according to claim 1, characterized in that, The edge feature perception in step (4) includes: (4-1) Two-dimensional discrete cosine transform is performed on the original camouflage image to extract frequency domain representation, and high frequency components are intercepted; after size normalization through depth separable convolution and upsampling, the channel is compressed and the frequency attention map is obtained through Sigmoid; (4-2) The coarse mask is generated into foreground attention map through bilinear interpolation and Sigmoid, and the background attention map is obtained by taking the inverse; the spatial enhancement features are multiplied element by element with the frequency, foreground and background three attention maps, and then spliced along the channel; after passing through the convolution module, batch normalization, ReLU and Sigmoid in turn, the edge attention is generated, and then the original spatial enhancement features are added in residual form to obtain the edge features; (4-3) The edge features are reduced to single-channel edge probability map through convolution, and the real edge mask is used as a supervision signal to complete the training of the edge detection branch.

6. The camouflage target detection method based on feature selection attention and frequency domain edge guidance according to claim 1, characterized in that, The skip connection decoding in step (5) includes: (5-1) The top layer spatial enhancement features are obtained through the "convolution-batch normalization-ReLU" one-level decoder to obtain the top layer decoding features, which are sent into the two-level decoder together with the next layer edge features to obtain the two-level decoding features; the two-level decoding features are spliced along the channel and iterated layer by layer downward according to the same strategy, and finally four layers of decoding features are output; (5-2) Each layer of decoding features is reduced to a single-channel mask prediction map through convolution, and is connected in residual with the next layer prediction map; each layer of prediction map is supervised at the pixel level by the real mask, and finally all layers of prediction map are added element by element to generate the final camouflage target mask.

Images