Camouflage target detection method based on edge semantic collaboration

Abstract

The invention discloses a camouflage target detection method based on edge semantic collaboration, and the method comprises the steps: an encoder which extracts multi-scale features from an original input image; the decoder comprises an adaptive edge texture perceptron AETP, a double-flow feature enhancer DSFA and a multi-feature modulation module MFMM; an adaptive edge texture perceptron AETP receives the multi-scale features, and extracts edge features by using multi-scale deformable convolution fusion and cross attention guidance; a double-flow feature enhancer DSFA performs feature enhancement on the multi-scale feature based on the edge feature to obtain a multi-scale enhanced feature; a multi-feature modulation module MFMM modulates the enhanced features of each scale based on the edge information to obtain a prediction image; the last level of prediction image is a camouflage target detection result. According to the method, a dynamic coupling edge-texture perception method is adopted, so that the capability of detecting a camouflage object in a challenging scene by the model is greatly improved.

Description

Technical Field

[0001] The present invention relates to the field of visual detection technology, and in particular to a camouflaged target detection method based on edge semantic collaboration. Background Art

[0002] Camouflaged object detection has become a significant challenge in computer vision, aiming to identify objects that blend seamlessly into their surroundings. Early methods relied on hand-crafted features, such as texture contrast and color contrast, to segment camouflaged objects. However, due to their limited adaptability, these methods struggled in complex scenes. The advent of large-scale datasets and deep learning has enabled significant progress in camouflaged object detection. Modern approaches can be broadly categorized into three paradigms: 1) Advanced network architectures: Many works focus on enhancing feature representations through specialized modules. For example, SINet introduces a biologically inspired network that mimics the hunting behavior of predators to iteratively search for and localize camouflaged objects. Transformer-based models such as DTINet and FSPNet utilize self-attention to model long-range dependencies, while Bayesian learning is integrated into the Transformer to combine deterministic and probabilistic reasoning. Multi-scale aggregation strategies such as ZoomNet, OAformer, and CubeNet utilize attention fusion and X-shaped connections to unify features across different scales. Furthermore, a growing number of models are using diffusion to learn denoising processes for object localization and detection. 2) Multi-task and multi-stage learning: Joint frameworks often incorporate auxiliary tasks to facilitate segmentation. LSR and DGNet integrate ranking, localization, and gradient generation with COD. Multi-stage methods such as SINet-V2 and SegMaR adopt a coarse-to-fine pipeline, first localizing the target and then refining the details through zooming in or iterative learning. PFNet combines location and focus modules with attention mining to mimic the human recognition process. 3) Biologically inspired strategies: Inspired by natural predation or human vision, methods such as PreyNet divide detection into initial localization and predator-like refinement. Similarly, SINet's search recognition network simulates the dynamic relationship between predators and prey. Rank-Net draws inspiration from the biological characteristics of animal camouflage mechanisms and establishes a hierarchical understanding of object hiding. The architecture proposed by DAT draws inspiration from biological vision by synergistically combining local and global contextual cues through dual aggregate transformers, enhancing the super-resolution of images and restoring the fine structural details necessary for downstream detection tasks.

[0003] Traditional segmentation methods rely on color and texture contrast, which fail in complex camouflaged scenes. Existing deep learning methods that rely on isolated edge detection or unimodal texture analysis often produce fragmented predictions due to three fundamental limitations: 1) the fixed geometric priors in traditional convolution cannot adapt to irregular boundaries and random textures; 2) the separate processing of edge and texture features hinders mutual reinforcement; and 3) there is a lack of a hierarchical calibration mechanism for multi-scale edge-texture consistency. Summary of the Invention

[0004] The purpose of the present invention is to solve the problems in the prior art.

[0005] The technical solution adopted by the present invention to solve the technical problem is: to provide a camouflaged target detection method based on edge semantic collaboration, using an edge semantic collaboration network model, the edge semantic collaboration network model includes:

[0006] Encoder, extracts multi-scale features from the original input image;

[0007] Adaptive Edge Texture Perceptor (AETP) receives the multi-scale features output by the encoder and extracts edge features using multi-scale deformable convolution fusion and cross-attention guidance;

[0008] The dual-stream feature enhancer DSFA enhances multi-scale features based on edge features to obtain multi-scale enhanced features;

[0009] The multi-feature modulation module MFMM modulates the enhanced features at each scale to obtain a predicted image; the predicted image at the last level is the camouflaged target detection result.

[0010] Preferably, the process of extracting multi-scale features from the original input image by the encoder comprises the following steps:

[0011] PVT-v2 is used to extract multi-scale pyramid features from a given original input image I, which is expressed as Enc means using the pre-trained PVT-v2 encoder;

[0012] The multi-scale pyramid features are aggregated and input into the void space pyramid pooling module ASPP to generate the final feature map F′4=A e [F1,F2,F3,F4]; A e represents the void space pyramid pooling module ASPP;

[0013] Encoder output and F′4 as multi-scale features.

[0014] Preferably, the adaptive edge texture perceptron AETP receives multi-scale features output by the encoder and extracts edge features using multi-scale deformable convolution fusion and cross attention guidance, comprising the following steps:

[0015] Use deformable convolutions with texture-conditional offsets to build adaptive edge representations:

[0016]

[0017] in, Indicates the first convolution, batch normalization, activation function and second convolution operations are performed in sequence; express The output of the operation; Represents the operations of deformable convolution, batch normalization, activation function and convolution in sequence, θ offset It is the offset generated by a specialized convolution branch inside the deformable convolution using the input features; Indicates a dedicated convolution branch; ↑ indicates an upsampling operation; Represents element-by-element addition operation; f i dc Indicates the corresponding f i Adaptive edge representation of ;

[0018] Apply the multi-head self-attention mechanism for calculation, and then gradually concatenate the calculation results to obtain the edge features, which can be expressed as:

[0019]

[0020] in, Represents the self-attention mechanism, [] represents the channel dimension connection, and s1 and s2 both represent the calculation results of self-attention; It represents convolution, batch normalization, activation function and Sigmoid function in sequence; E represents the predicted edge map.

[0021] Preferably, the dual-stream feature enhancer DSFA performs feature enhancement on multi-scale features based on edge features to obtain multi-scale enhanced features, comprising the following steps:

[0022] The original input image is adaptively cropped and fused with the extracted features to generate a pixel patch reference, which is expressed as:

[0023]

[0024] Where I represents the original input image; PatchRef represents the adaptive cropping reference operation, specifically refers to cropping I into F i Several patches of the same size; represents the pixel patch reference, k represents the kth patch, and N represents the number of patches after adaptive cropping; represents the convolution operation, dw represents the pixel patch feature; Represents the operations of deformable convolution, batch normalization, activation function and convolution in sequence; Δp represents the offset learned in the deformable convolution layer, which is used to guide the adaptive sampling position of the deformable convolution and strengthen the geometric and texture alignment of key boundaries;

[0025] Texture enhancement of features is performed using pixel reference patches, expressed as:

[0026]

[0027] in, represents the enhanced features of the i-th scale, y1, y2, y3 and y4 represent texture enhanced features; It means that the convolution, batch normalization and activation function operations of the x×x size convolution kernel are performed in sequence.

[0028] Preferably, the multi-feature modulation module MFMM modulates the enhanced features of each scale based on edge information to obtain a predicted image, comprising the following steps:

[0029] Receive input features Expressed as:

[0030]

[0031] in, represents the enhanced features of the i-th scale, represents the output features of the multi-feature modulation module MFMM at level i;

[0032] Use three branches to perform attention calculation on the enhanced features respectively, and connect the three calculation results to obtain the connection result;

[0033] The connection result is modulated by channel attention and spatial attention to obtain the predicted image.

[0034] Preferably, the method of using three branches to respectively perform attention calculation on the enhanced features and performing a connection operation on the three calculation results to obtain a connection result comprises the following steps:

[0035] Receive enhanced features, edge features of the current scale and the predicted image of the previous scale;

[0036] The mask-guided branch performs attention calculation on the enhanced features of the current scale based on the predicted image of the previous scale; the edge-guided branch performs attention calculation on the enhanced features of the current scale based on the edge features; the Laplace-guided grouping performs attention calculation on the enhanced features of the current scale based on the Laplace operator, which can be expressed as:

[0037]

[0038] Among them, the enhanced feature is used as the input feature F for attention calculation in , and Represents three different convolution operations; Γ ref. represents the conditioning texture used by the three guiding branches, Represents the calculation results of the three guiding branch outputs, ref.={Mask,Edge,Lp};ref.=Mask,Γ Mask Indicates mask information, Represents the calculation result of the mask-guided branch output; when ref.=Edge, Γ Edge Represents edge information, represents the calculation result of the edge-guided branch output; when ref.=Lp, Γ Lp represents the Laplace operator information, represents the calculation result of the Laplace guided branch output;

[0039] The attention calculation results output by the three branches are connected to obtain a connection result, which is expressed as:

[0040]

[0041] Preferably, the connection result is modulated by channel attention and spatial attention to obtain a predicted image, comprising the following steps:

[0042] The convolutional block attention is used to enhance channel and spatial attention, which is expressed as:

[0043]

[0044] in, represents the output of the i-th level MFMM; CBAM represents the convolutional block attention module, which mainly consists of channel attention and spatial attention; M i Represents the predicted image of level i; [] represents the channel dimension connection operation; Indicates that convolution, batch normalization, activation function and Sigmoid function are performed in sequence;

[0045] Preferably, the edge semantic collaborative network model adopts three loss functions, including binary cross entropy loss, intersection-over-union loss and Dice loss; binary cross entropy loss and intersection-over-union loss are used to simultaneously supervise the predicted mask, and Dice loss is used to supervise the predicted edge;

[0046] Binary cross entropy loss The pixel-level classification supervision for predicted masks and true masks is expressed as:

[0047]

[0048] where y n ∈{0,1} is the true label, p n ∈(0,1) is the prediction probability, M is the total number of pixels;

[0049] Intersection-over-Merge Loss Supervision is performed on the shape consistency of the mask, which is expressed as:

[0050]

[0051] Among them, P is the predicted mask and Y is the true label;

[0052] Days Loss The degree of matching between the predicted edge and the true edge is supervised and expressed as follows:

[0053]

[0054] Here, ∈ is a smoothing term.

[0055] The present invention has the following beneficial effects:

[0056] (1) The use of dynamically coupled edge-texture perception in the present invention greatly improves the model's ability to detect camouflaged objects in challenging scenarios;

[0057] (2) In the multi-feature modulation module of the present invention, the mask feature itself has a strong baseline performance, the edge guidance helps to recover the boundary details lost in the mask prediction, and the Laplacian operator improves the sensitivity to fine texture transformation.

[0058] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments, but the present invention is not limited to the embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 This is a framework diagram of the edge semantic collaborative network model according to an embodiment of the present invention;

[0060] Figure 2 This is a network structure diagram of an adaptive edge texture sensor according to an embodiment of the present invention;

[0061] Figure 3 1 is a network structure diagram of a dual-stream feature enhancer according to an embodiment of the present invention;

[0062] Figure 4 This is a network structure diagram of a multi-feature modulation module according to an embodiment of the present invention;

[0063] Figure 5 Schematic diagram showing the visual differences in detection effects between the embodiment of the present invention and other methods. DETAILED DESCRIPTION

[0064] See also Figure 1 The following is a schematic diagram of the framework structure of the edge semantic collaborative network model of an embodiment of the present invention, which consists of an encoder and multiple collaborative components to gradually decode camouflaged objects. The architecture adopts a layered refinement method, including:

[0065] Encoder, using PVT-v2 as encoder Enc. Given the original input image The encoder extracts multi-scale pyramid features, represented as These features can effectively capture global patterns and local deceptive textures, which are crucial for the recognition of camouflaged objects. In addition, to further enhance the receptive field and integrate multi-scale semantic information, the output of each layer of the encoder is aggregated and input into the Atrous Spatial Pyramid Pooling (ASPP) module A. e , and finally generate the final feature map f′4=A e [F1,F2,F3,F4].

[0066] Other components: The proposed decoder framework consists of three key modules: adaptive edge texture perceptron (AETP), dual-stream feature enhancer (DSFA), and multi-feature modulation module (MFMM); the architecture hierarchically integrates these modules to achieve dynamic edge texture coupling and progressive prediction improvement.

[0067] See also Figure 2 The figure shows the network structure of the adaptive edge texture perceptron (AETP) in an embodiment of the present invention. Through multi-scale deformable fusion and cross-attention guided refinement, it overcomes the fixed geometric prior in the traditional COD method and establishes a symbiotic edge texture perception. This module processes the multi-scale features output by the encoder through two key operations. and F′4:

[0068] Texture-aware deformable fusion, using deformable convolutions with texture-conditional offsets to build adaptive edge representations:

[0069]

[0070] in Represents convolution, batch normalization (BN), activation function (ReLU), and convolution. represents the deformable convolution layer, θ offset It is generated by a dedicated convolution branch;

[0071] Cross-scale attention guidance, which applies a multi-head self-attention mechanism that enforces edge continuity by focusing on local features, allowing the model to emphasize relevant parts of the image while suppressing irrelevant background information. The self-attention operation is designed to capture long-range dependencies and enhance the discriminability of features, which is crucial for identifying subtle differences between camouflaged and non-camouflaged areas;

[0072] Finally, an edge prediction method is formed by gradually connecting and improving edge features.

[0073]

[0074] in, represents the self-attention mechanism, and […] represents the channel-wise connection. Represents convolutional layers, BN, ReLU, and Sigmoid. The combination of deformable convolution and self-attention enables the model to effectively handle irregular and blurred edges, thereby improving the overall edge prediction effect.

[0075] Since directly using edge information from AETP to guide encoder features is not effective, the present invention proposes a dual-stream feature enhancer, see Figure 3 As shown in Figure 3, the dual-stream feature enhancer DSFA utilizes edge-aware geometric features and multi-scale contextual information to achieve enhancement. The enhancement consists of three stages: pixel patch reference, geometric texture fusion, and feature texture enhancement.

[0076] Pixel Patch Reference: Disguised objects have a high degree of similarity to their backgrounds, so subtle differences in texture, edges, and context are crucial for accurate recognition. We propose Pixel Patch Reference, which adaptively crops the original input image and fuses it with the extracted features to provide precise reference information. Unlike traditional downsampling methods, Pixel Patch Reference ensures that the cropped patches are spatially consistent with the features, thus compensating for information loss in deep networks:

[0077]

[0078] Where I represents the original input image, Represents a high-resolution pixel patch reference.

[0079] Geometric texture fusion, in order to effectively capture the transformation of boundaries and textures, edge-conditional deformation convolution is performed on the merged features:

[0080]

[0081] Among them, F i and E are the encoder features and edge map respectively; Representing the offset learned by the deformable convolutional layer, Δp guides the adaptive sampling and enforces the geometric and texture alignment of key boundaries.

[0082] Feature texture enhancement, the cascaded residual architecture can gradually expand the receptive field, thereby enriching features of different scales, which can be expressed as:

[0083]

[0084]

[0085] Among them, y i Represents the output of the i-th scale. is the output of the i-th DSFA. This structure ensures that each scale contains the deformed features dw and the refined output of the previous stage, thus achieving gradual and effective feature enhancement.

[0086] See also Figure 4 Figure 2 shows the network architecture of the Multi-Feature Modulation Module (MFMM) in an embodiment of the present invention. Due to the large scale variations and high background similarity of camouflaged objects, accurate mask prediction, especially edge prediction, presents significant challenges. To address this issue, the MFMM implements hierarchical edge-texture mutual verification to address the prediction fragmentation caused by the deceptive textures of camouflaged objects and gradually optimizes prediction accuracy, particularly edge prediction accuracy. This operates through two interrelated mechanisms: structurally constrained multi-texture attention (SCA) and hierarchical feature calibration.

[0087] The structure-constrained multi-texture attention mechanism leverages the edge priors obtained from the previous modules and extends the masked attention through multi-branch feature conditioning with geometric priors. For each branch, a structure-constrained self-attention layer is used to focus the model on key areas. In addition, three complementary priors are added to effectively guide the attention process:

[0088] (1) Mask guidance, injecting coarse spatial activation M from upper layer prediction i+1 , dynamically focusing computation on regions that are likely to contain the target object. This mimics the biological attention mechanism and suppresses irrelevant background textures.

[0089] (2) Edge guidance, which integrates boundary features from the AETP module through element-wise multiplication with the key / value. This hardwires geometric constraints into the attention process, preventing boundary drift caused by texture illusions.

[0090] (3) Laplacian guidance: Laplacian mapping is used to enhance high-frequency details and offset the texture blur in deep CNN features. It amplifies high-frequency details that may be suppressed in the original features and promotes the recovery of fine textures and structures.

[0091] The conditional attention formula of each branch is expressed as:

[0092]

[0093] in, represents depth convolution, F represents and Γ ref. Represents three adjustment textures, ref.={Mask,Edge,Lp}; ref.=Mask,Γ Mask Indicates mask information, Represents the calculation result of the mask-guided branch output; when ref.=Edge, Γ Edge Represents edge information, represents the calculation result of the edge-guided branch output; when ref.=Lp, Γ Lp represents the Laplace operator information, represents the calculation result of the Laplace guided branch output;

[0094] Hierarchical feature calibration, multi-branch outputs are aggregated through channel connection and spatial refinement, using the connection operation to integrate information from multiple branches. A convolutional block attention module (CBAM) is then added to enhance channel and spatial attention. Channel attention compresses redundant texture representations, while spatial attention makes boundary transitions clearer. Finally, a convolution operation is applied to obtain the predicted image M from the MFMM. i The specific formula is as follows:

[0095]

[0096] Among them, [] represents the connection operation, Represents convolution, BN, and ReLU activation.

[0097] At the same time, the input features corresponding to each MFMM Equal to the output of the corresponding DSFA module And the output of the previous layer MFMM The summation operation is expressed as In particular, In addition, the smallest features are obtained directly from the convolutional layer, and the corresponding output is

[0098] Specifically, the edge semantic collaborative network model is trained using three loss functions, including binary cross entropy loss, intersection-over-union loss, and Dice loss. Binary cross entropy loss and intersection-over-union loss are used to simultaneously supervise the predicted mask, while Dice loss is used to supervise the predicted edge.

[0099] Binary cross entropy loss The pixel-level classification supervision for predicted masks and true masks is expressed as:

[0100]

[0101] where y n ∈{0,1} is the true label, p n ∈(0,1) is the prediction probability, M is the total number of pixels;

[0102] Intersection-over-Merge Loss Supervision is performed on the shape consistency of the mask, which is expressed as:

[0103]

[0104] Among them, P is the predicted mask and Y is the true label;

[0105] Days Loss The degree of matching between the predicted edge and the true edge is supervised and expressed as follows:

[0106]

[0107] Here, ∈ is a smoothing term.

[0108] The embodiment of the present invention is verified by experiments, and four widely used evaluation indicators are used: mean absolute error (M), structural metric (S m ), weighted F-measure and the average E metric (E φ ). The mean absolute error measures the absolute difference between the predicted results and the true values, providing a direct quantification of the prediction accuracy. The structural metric simultaneously evaluates the region-aware and object-aware structural similarity between the predicted results and the true values, which is crucial for evaluating how well the model captures the boundaries and internal structures of camouflaged objects. The weighted F-measure combines recall and precision to provide a comprehensive evaluation of detection performance. Finally, the average E-measure calculates element-wise similarity and aggregates it at the image level, which helps capture fine-grained performance. The experimental results are shown in Figure 2. Figure 1 As shown in Tables 1 to 3.

[0109] Table 1 - Quantitative comparison results of the method of the embodiment of the present invention and other methods on different benchmark data sets:

[0110]

[0111] The datasets listed in Table 1 include NC4K, COD10K, and CAMO. As can be seen from Table 1, the method of the embodiment of the present invention consistently outperforms existing methods in terms of mean absolute error and mean E-metric, and the use of dynamically coupled edge-texture perception greatly improves the model's ability to detect camouflaged objects in challenging scenes. For a visual comparison of the detection results between the method of the embodiment of the present invention and some other methods, see Figure 5 As shown, it can be seen that the method of the embodiment of the present invention has obvious advantages over other methods in predicting the edges of camouflaged target details, especially in areas with dense edges.

[0112] Table 2 - Ablation test results of three modules in the embodiment of the present invention:

[0113]

[0114] The three modules described in Table 2 refer to the multi-feature modulation module, the adaptive edge texture sensor, and the dual-stream feature enhancer. As can be seen from Table 2, after adding the multi-feature modulation module, the adaptive edge texture sensor, and the dual-stream feature enhancer in sequence, the experimental results have steadily improved. Among them, the baseline refers to the result obtained using the encoder and a simple convolutional layer.

[0115] Table 3 - Ablation experiment of the multi-feature modulation module of the embodiment of the present invention:

[0116]

[0117] As can be seen from Table 3, the mask feature itself has a strong baseline performance, the edge guidance helps to recover the boundary details lost in the mask prediction, and the Laplacian operator improves the sensitivity to fine texture transformations.

[0118] As can be seen, the present invention's dynamically coupled edge-texture perception method significantly improves the model's ability to detect camouflaged objects in challenging scenarios. In the multi-feature modulation module, the mask feature itself has strong baseline performance, edge guidance helps restore boundary details lost in mask prediction, and the Laplacian operator improves sensitivity to fine texture transitions. This significantly improves the ability to detect camouflaged objects.

[0119] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A camouflaged target detection method based on edge semantic collaboration, characterized in that: An edge semantic collaborative network model is adopted, and the edge semantic collaborative network model includes: Encoder, extracts multi-scale features from the original input image; Adaptive Edge Texture Perceptor (AETP) receives the multi-scale features output by the encoder and extracts edge features using multi-scale deformable convolution fusion and cross-attention guidance; The dual-stream feature enhancer DSFA enhances multi-scale features based on edge features to obtain multi-scale enhanced features; The multi-feature modulation module MFMM modulates the enhanced features at each scale to obtain a predicted image; the predicted image at the last level is the camouflaged target detection result.

2. The method for detecting disguised targets based on edge semantic collaboration according to claim 1, characterized in that: The process of extracting multi-scale features from the original input image by the encoder includes the following steps: PVT-v2 is used to extract multi-scale pyramid features from a given original input image I, which is expressed as Enc means using the pre-trained PVT-v2 encoder; The multi-scale pyramid features are aggregated and input into the void space pyramid pooling module ASPP to generate the final feature map F4′=A e [F1,F2,F3,F4]; A e represents the void space pyramid pooling module ASPP; Encoder output and F4′ as multi-scale features.

3. The method for detecting camouflaged targets based on edge semantic collaboration according to claim 1, wherein: The adaptive edge texture perceptron (AETP) receives the multi-scale features output by the encoder and extracts edge features using multi-scale deformable convolution fusion and cross attention guidance, including the following steps: Use deformable convolutions with texture-conditional offsets to build adaptive edge representations: in, Indicates the first convolution, batch normalization, activation function and second convolution operations are performed in sequence; express The output of the operation; Represents the operations of deformable convolution, batch normalization, activation function and convolution in sequence, θ offset It is the offset generated by a specialized convolution branch inside the deformable convolution using the input features; Indicates a dedicated convolution branch; ↑ indicates an upsampling operation; Represents element-by-element addition operation; f i dc Indicates the corresponding f i Adaptive edge representation of ; Apply the multi-head self-attention mechanism for calculation, and then gradually concatenate the calculation results to obtain the edge features, which can be expressed as: in, Represents the self-attention mechanism, [] represents the channel dimension connection, and s1 and s2 both represent the calculation results of self-attention; It represents convolution, batch normalization, activation function and Sigmoid function in sequence; E represents the predicted edge map.

4. The method for detecting camouflaged targets based on edge semantic collaboration according to claim 1, characterized in that: The dual-stream feature enhancer DSFA performs feature enhancement on multi-scale features based on edge features to obtain multi-scale enhanced features, including the following steps: The original input image is adaptively cropped and fused with the extracted features to generate a pixel patch reference, which is expressed as: Where I represents the original input image; PatchRef represents the adaptive cropping reference operation, specifically refers to cropping I into F i Several patches of the same size; represents the pixel patch reference, k represents the kth patch, and N represents the number of patches after adaptive cropping; represents the convolution operation, dw represents the pixel patch feature; Represents the operations of deformable convolution, batch normalization, activation function and convolution in sequence; Δp represents the offset learned in the deformable convolution layer, which is used to guide the adaptive sampling position of the deformable convolution and strengthen the geometric and texture alignment of key boundaries; Texture enhancement of features is performed using pixel reference patches, expressed as: in, represents the enhanced features of the i-th scale, y1, y2, y3 and y4 represent texture enhanced features; It means that the convolution, batch normalization and activation function operations of the x×x size convolution kernel are performed in sequence.

5. The method for detecting camouflaged targets based on edge semantic collaboration according to claim 1, wherein: The multi-feature modulation module MFMM modulates the enhanced features of each scale based on edge information to obtain a predicted image, including the following steps: Receive input features Expressed as: in, represents the enhanced features of the i-th scale, represents the output features of the multi-feature modulation module MFMM at level i; Use three branches to perform attention calculation on the enhanced features respectively, and connect the three calculation results to obtain the connection result; The connection result is modulated by channel attention and spatial attention to obtain the predicted image.

6. The method for detecting camouflaged targets based on edge semantic collaboration according to claim 5, characterized in that: The method of using three branches to respectively perform attention calculation on the enhanced features and performing a connection operation on the three calculation results to obtain a connection result includes the following steps: Receive enhanced features, edge features of the current scale and the predicted image of the previous scale; The mask-guided branch performs attention calculation on the enhanced features of the current scale based on the predicted image of the previous scale; the edge-guided branch performs attention calculation on the enhanced features of the current scale based on the edge features; the Laplace-guided grouping performs attention calculation on the enhanced features of the current scale based on the Laplace operator, which can be expressed as: Among them, the enhanced feature is used as the input feature F for attention calculation in , and Represents three different convolution operations; Γ ref. represents the conditioning texture used by the three guiding branches, Represents the calculation results of the three guiding branch outputs, ref.={Mask,Edge,Lp};ref.=Mask,Γ Mask Indicates mask information, Represents the calculation result of the mask-guided branch output; when ref.=Edge, Γ Edge Represents edge information, represents the calculation result of the edge-guided branch output; when ref.=Lp, Γ Lp represents the Laplace operator information, represents the calculation result of the Laplace guided branch output; The attention calculation results output by the three branches are connected to obtain a connection result, which is expressed as:

7. The method for detecting disguised targets based on edge semantic collaboration according to claim 1, characterized in that: The connection result is modulated by channel attention and spatial attention to obtain a predicted image, including the following steps: The convolutional block attention is used to enhance channel and spatial attention, which is expressed as: in, represents the output of the i-th level MFMM; CBAM represents the convolutional block attention module, which mainly consists of channel attention and spatial attention; M i Represents the predicted image of level i; [] represents the channel dimension connection operation; It means that convolution, batch normalization, activation function and Sigmoid function are performed in sequence.

8. The method for detecting disguised targets based on edge semantic collaboration according to claim 1, characterized in that: The edge semantic collaborative network model adopts three loss functions, including binary cross entropy loss, intersection-over-union loss, and Dice loss. Binary cross entropy loss and intersection-over-union loss are used to simultaneously supervise the predicted mask, while Dice loss is used to supervise the predicted edge. Binary cross entropy loss The pixel-level classification supervision for predicted masks and true masks is expressed as: where y n ∈{0,1} is the true label, p n ∈(0,1) is the prediction probability, M is the total number of pixels; Intersection-over-Merge Loss Supervision is performed on the shape consistency of the mask, which is expressed as: Among them, P is the predicted mask and Y is the true label; Days Loss The degree of matching between the predicted edge and the true edge is supervised and expressed as follows: Here, ∈ is a smoothing term.

Images