Language-guided structure-aware network architecture and method for camouflage target detection

Abstract

The invention discloses a language-guided structure-aware network architecture and method for camouflage target detection, and relates to the technical field of camouflage target detection. According to the method, the text is guided to focus on a potential target, text semantic and visual features are fused through the CLIP model, the target mask is generated, the problems that a traditional model lacks semantic guidance and is difficult to focus on a camouflage area are solved, and background interference is greatly reduced; edge details can be accurately extracted, a Fourier edge enhancement module (FEEM) is combined with spatial domain edge enhancement and frequency domain high-frequency information capture, the problem of a fuzzy pain point of a camouflage target boundary is effectively solved, and the edge positioning precision is improved; structural and local dual optimization is proposed, a structural awareness attention module (SAAM) is fused with semantic and edge information, and a coarse guidance local refinement module (CGLRM) is designed through double branches of global guidance and local refinement, so that the regional consistency and structural integrity of a segmentation result are guaranteed, and local details are prevented from being lost.

Description

Technical Field

[0001] This invention relates to the field of camouflage target detection technology, specifically to a language-guided structure-aware network architecture and method for camouflage target detection. Background Technology

[0002] Camouflaged target detection is an important and challenging task in computer vision. Its core objective is to identify objects in complex backgrounds that are highly similar to their surroundings in terms of color, texture, and shape. Because camouflaged targets inherently lack saliency, they often exhibit features such as blurred boundaries, weakened contours, and extremely low semantic responses, making it difficult for traditional semantic segmentation methods to achieve satisfactory performance on such tasks.

[0003] Camouflaged target detection has significant applications in numerous fields, such as natural ecological monitoring (e.g., wildlife conservation), medical image analysis (e.g., lesion detection), industrial defect inspection, and military concealed target detection. Therefore, developing efficient and robust camouflaged target detection methods has important theoretical and practical value. Compared to general target detection (GOD) and salient target detection (SOD), the main challenge of COD lies in the high similarity of visual features between the target and the background, which limits the effectiveness of traditional methods based on saliency or contrast.

[0004] In recent years, the camouflaged target detection (COD) task still faces many challenges. First, COD models usually rely on a single visual backbone network and lack explicit guidance from textual semantic priors, making it difficult to effectively focus on camouflaged regions. Second, camouflaged targets often exhibit blurred boundaries and weakened contours, making accurate edge localization difficult. Third, the internal structure of camouflaged regions is both complex and subtle, making fine-grained local modeling difficult, which in turn affects the regional consistency and structural integrity of the segmentation results.

[0005] Therefore, a new solution is needed to address the above problems. Summary of the Invention

[0006] The purpose of this invention is to provide a language-guided structure-aware network architecture and method for detecting camouflaged targets, in order to solve the technical problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a language-guided structure-aware method for camouflaged target detection, comprising at least the following steps:

[0008] The visual backbone is a PVT-v2 network, using PVT-v2-b3 as the image feature extractor to extract features from the input image I∈R. H×W×3 The four-scale feature map is obtained, denoted as... ;

[0009] To enhance the semantic information of the camouflaged target, a frozen CLIP model is further introduced. The text encoder of the CLIP model extracts text features from text cues (such as "A photo of the camouflaged owl"), while the visual encoder extracts multi-scale visual features from layers 8, 16, and 24.

[0010] Cross-scale alignment and fusion are performed using a Feature Pyramid Network (FPN) to obtain CLIP visual features Attn. CLIP visual features Attn and text features are then input into Text-Guided Attention to generate a target mask. ;

[0011] mask Used via MGFA operation The explicit feature enhancement of ·Attn produces and This allows the model to focus on potential camouflage areas;

[0012] Enhanced features First, input the Transformer to obtain the attn features;

[0013] The Fourier edge enhancement module uses Using attn as input, generate edge-guided features E;

[0014] In the next stage, the highest-level features will be... The attn features are concatenated and input into the coarse-guided local refinement module to obtain coarse-grained structural features. The coarse-grained structural features, along with E and attn features, are input into the structure-aware attention module to obtain structural enhancement features.

[0015] Then, it is further refined through a local refinement module guided by coarse-grained principles, generating... ;

[0016] For the remaining scales i=3,2,1, the refinement process is as follows:

[0017] First, output the previous scale. Upsampling and comparing with the corresponding backbone features cascade;

[0018] Then compressed into convolutional blocks ;

[0019] Will Edge-guided features E and The features are fed into the structure-aware attention module, and then the output features of the structure-aware attention module are locally refined through a coarse-guided local refinement module to obtain the output. ;

[0020] During the prediction phase, the network... Convolution outputs segmentation results at multiple scales, by and Intermediate prediction plots were obtained respectively. and And upsampled to the original resolution by interpolation;

[0021] semantic-guided prediction generated by attn Edge prediction is obtained from E. This also includes a mask. ;

[0022] The final output set is: O={ , , , , , }

[0023] Furthermore, the operation of the Fourier edge enhancement module includes at least the following steps:

[0024] Based on features and attn as input;

[0025] First, the highest layer features Concatenate with attn and obtain high-level fused features through channel compression convolution:

[0026]

[0027] Subsequently, this feature is upsampled stepwise and concatenated with lower-level features, then fused via convolution, and the boundary representation is enhanced using the EdgeEnhancer module.

[0028]

[0029]

[0030]

[0031] in, This represents the concatenation operation of features. Indicates feature upsampling, This refers to the EdgeEnhancer module. Represents convolution;

[0032] The EdgeEnhancer module is based on the idea of ​​average pooling difference:

[0033] First, smooth the input features, then calculate their difference to extract high gradient regions (i.e. edges).

[0034] An edge weight map is then generated and fused with the original features using a residual approach, thereby explicitly enhancing boundary contrast in the spatial domain. Its mathematical form is as follows:

[0035]

[0036]

[0037]

[0038] Where Pool represents a 1×1 average pooling operation, and σ represents the Sigmoid function. Indicates batch normalization, Represents a 1×1 convolution. Represents the edge weight graph;

[0039] To further extract high-frequency detail information of the target boundary, in the fused features Applying Fourier transform to explicitly capture the high-frequency response:

[0040]

[0041] in represent function Represents the Fourier transform;

[0042] Finally, output edge-guided features: .

[0043] Furthermore, the application of the structure-aware attention module includes at least the following steps:

[0044] Given input features x ,Right now ;

[0045] First, the Q, K, and V features are extracted from feature x through convolution;

[0046] Then, feature M is applied to Q (emphasizing the target structure), and edge-guided feature E is applied to K (highlighting boundary information);

[0047]

[0048]

[0049]

[0050] in, Represents convolution; This indicates element-wise multiplication;

[0051] Then Q, K, Transform the token representation from [B,C,H,W] into [B,HW,C], and then proceed. Vector normalization;

[0052] To further reduce memory usage, approximate attention computation is used:

[0053] First, perform Softmax on the transposed K, then... Multiply, and finally sum Multiply;

[0054]

[0055]

[0056]

[0057] in, This represents batch matrix multiplication;

[0058] Finally, the output is reshaped to recover [B,C,H,W] and fused with the original input features via residual connections.

[0059] Furthermore, the application of the coarse-guided local refinement module includes at least the following steps:

[0060] Given an input feature map x∈ This invention first applies channel attention (CA) and spatial attention (SA) to extract global coarse guiding features:

[0061]

[0062]

[0063] in, This indicates element-wise multiplication;

[0064] Attention-enhancing feature G provides coarse-grained spatial guidance for the target;

[0065] Next, the input feature x is spatially divided into four non-overlapping sub-regions:

[0066]

[0067] The attention-enhancing feature G was divided in the same way as This is to ensure that each local area is equipped with the corresponding global guidance information;

[0068]

[0069] Where σ(·) represents the Sigmoid activation used to generate the smooth guiding weights. Indicates the convolution operation;

[0070] All the local refinement results are stitched together into a complete feature map:

[0071]

[0072] Subsequently, with attention enhancement features To merge:

[0073]

[0074] Finally, the output enhancement module generates the final output:

[0075]

[0076] in, Indicates the convolution operation; This indicates element-wise addition; This represents the splicing operation of features.

[0077] A language-guided structure-aware network architecture for camouflaged target detection, comprising a PVT-v2 backbone network, a CLIP model, a Fourier edge enhancement module (FEEM), a structure-aware attention module (SAAM), and a coarse-guided local refinement module (CGLRM).

[0078] The text prompts and RGB images are input into the CLIP model to obtain the text and visual features of the disguised target. The text-guided attention is used to generate a target mask, which guides the multi-scale features output by the PVT-v2 backbone network to focus on the potential target region.

[0079] The Fourier Edge Enhancement Module (FEEM) fuses multi-scale semantic features and extracts high-frequency information in the frequency domain to enhance edges. The resulting edge guidance features provide boundary cues for subsequent modules.

[0080] Furthermore, a Structure-Aware Attention Module (SAAM) is proposed to enhance the model's perception of target structures and boundaries;

[0081] Furthermore, a coarse-guided local refinement module (CGLRM) is proposed, employing a dual-branch structure:

[0082] One approach introduces channel attention and spatial attention to perform global perception modeling and generate attention-enhanced features;

[0083] The other branch divides the input features into four local regions;

[0084] Local structural refinement is performed under the guidance of attention-enhancing features.

[0085] Compared with the prior art, the beneficial effects of the present invention are:

[0086] 1. This invention guides the focus on potential targets by using text to integrate text semantics and visual features through the CLIP model to generate target masks. This solves the problems of traditional models lacking semantic guidance and having difficulty focusing on camouflaged areas, and significantly reduces background interference.

[0087] 2. This invention can accurately extract edge details. It uses a Fourier edge enhancement module (FEEM) to combine spatial domain edge enhancement with frequency domain high-frequency information capture, effectively solving the problem of blurred boundaries of camouflaged targets and improving edge positioning accuracy.

[0088] 3. This invention proposes dual optimization of structure and local aspects. The Structure-Aware Attention Module (SAAM) integrates semantic and edge information, while the Coarse-Guided Local Refinement Module (CGLRM) ensures regional consistency and structural integrity of the segmentation results through a dual-branch design of global guidance and local refinement, avoiding the loss of local details. Attached Figure Description

[0089] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0090] Figure 1 This is a general framework diagram of the present invention;

[0091] Figure 2 This is a diagram of the Fourier edge enhancement module of the present invention;

[0092] Figure 3 This is a diagram of the structural perception attention module of the present invention;

[0093] Figure 4 This is a diagram of the local refinement module of the coarse guidance of the present invention. Detailed Implementation

[0094] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0095] Example 1:

[0096] Please see Figure 1 A language-guided structure-aware method for camouflaged target detection includes at least the following steps:

[0097] The visual backbone is a PVT-v2 network, using PVT-v2-b3 as the image feature extractor to extract features from the input image I∈R. H×W×3 The four-scale feature map is obtained, denoted as... ;

[0098] To enhance the semantic information of the camouflaged target, a frozen CLIP model is further introduced. The text encoder of the CLIP model extracts text features from text cues (such as "A photo of the camouflaged owl"), while the visual encoder extracts multi-scale visual features from layers 8, 16, and 24.

[0099] Cross-scale alignment and fusion are performed using a Feature Pyramid Network (FPN) to obtain CLIP visual features Attn. CLIP visual features Attn and text features are then input into Text-Guided Attention to generate a target mask. ;

[0100] As shown in Figure 1, the mask Used via MGFA operation The explicit feature enhancement of ·Attn produces and This allows the model to focus on potential camouflage areas;

[0101] Enhanced features First, input the Transformer to obtain the attn features;

[0102] Fourier edge enhancement module with Using attn as input, generate edge-guided features E;

[0103] In the next stage, the highest-level features will be... The attn features are concatenated and input into the coarse-guided local refinement module to obtain coarse-grained structural features. The coarse-grained structural features, along with E and attn features, are input into the structure-aware attention module to obtain structural enhancement features.

[0104] Then, it is further refined through a local refinement module guided by coarse-grained principles, generating... ;

[0105] For the remaining scales i=3,2,1, the refinement process is as follows:

[0106] First, output the previous scale. Upsampling and comparing with the corresponding backbone features cascade;

[0107] Then compressed into convolutional blocks ;

[0108] Will Edge-guided features E and The features are fed into the structure-aware attention module, and then the output features of the structure-aware attention module are locally refined through a coarse-guided local refinement module to obtain the output. ;

[0109] During the prediction phase, the network... Convolution outputs segmentation results at multiple scales, by and Intermediate prediction plots were obtained respectively. and And upsampled to the original resolution by interpolation;

[0110] semantic-guided prediction generated by attn Edge prediction is obtained from E. This also includes a mask. ;

[0111] The final output set is: O={ , , , , , }

[0112] like Figure 2 As shown, the operation of the Fourier edge enhancement module includes at least the following steps:

[0113] Based on features and attn as input;

[0114] First, the highest layer features Concatenate with attn and obtain high-level fused features through channel compression convolution:

[0115]

[0116] Subsequently, this feature is upsampled stepwise and concatenated with lower-level features, then fused via convolution, and the boundary representation is enhanced using the EdgeEnhancer module.

[0117]

[0118]

[0119]

[0120] in, This represents the concatenation operation of features. Indicates feature upsampling, This refers to the EdgeEnhancer module. Represents convolution;

[0121] The EdgeEnhancer module is based on the idea of ​​average pooling difference:

[0122] First, smooth the input features, then calculate their difference to extract high gradient regions (i.e. edges).

[0123] An edge weight map is then generated and fused with the original features using a residual approach, thereby explicitly enhancing boundary contrast in the spatial domain. Its mathematical form is as follows:

[0124]

[0125]

[0126]

[0127] Where Pool represents a 1×1 average pooling operation, and σ represents the Sigmoid function. Indicates batch normalization, Represents a 1×1 convolution. Represents the edge weight graph;

[0128] To further extract high-frequency detail information of the target boundary, in the fused features Applying Fourier transform to explicitly capture the high-frequency response:

[0129]

[0130] in represent function Represents the Fourier transform;

[0131] Finally, output edge-guided features: .

[0132] like Figure 3 As shown, to enhance the model's ability to discriminate the boundary details and target structure of camouflaged targets, this invention proposes SAAM. Unlike traditional self-attention mechanisms that treat all spatial locations equally, SAAM incorporates the semantic information M (i.e., ...) of the camouflaged target. Figure 1 In (rich in global semantic information of the camouflaged target) and edge-guided features E (i.e. Figure 1The E in the model (rich in edge detail information of the camouflaged target) is introduced into attention modeling to achieve task-oriented guided attention.

[0133] The application of a structure-aware attention module includes at least the following steps:

[0134] Given input features x ,Right now ;

[0135] First, feature x is used to extract query Q, key K, and value V features through convolution;

[0136] Then, feature M is applied to the query vector (emphasizing the target structure), and edge-guided feature E is applied to the key vector (highlighting boundary information).

[0137]

[0138]

[0139]

[0140] in, Represents convolution; This indicates element-wise multiplication;

[0141] Then Q, K, Transform the token representation from [B,C,H,W] into [B,HW,C], and then proceed. Vector normalization;

[0142] To further reduce memory usage, approximate attention computation is used:

[0143] First, perform Softmax on the transposed K, then... Multiply, and finally sum Multiply;

[0144]

[0145]

[0146]

[0147] in, This represents batch matrix multiplication;

[0148] Finally, the output is reshaped to recover [B,C,H,W] and fused with the original input features via residual connections.

[0149] As shown in Figure 4, this invention proposes CGLRM. The application of the coarse-guided local refinement module includes at least the following steps:

[0150] Given an input feature map x∈ This invention first applies channel attention (CA) and spatial attention (SA) to extract global coarse guiding features:

[0151]

[0152]

[0153] in, This indicates element-wise multiplication;

[0154] Attention-enhancing feature G provides coarse-grained spatial guidance for the target;

[0155] Next, the input feature x is spatially divided into four non-overlapping sub-regions:

[0156]

[0157] The attention-enhancing feature G was divided in the same way as This ensures that each local area is equipped with corresponding global guidance information; this spatial division strategy effectively suppresses irrelevant interference by introducing global guidance into the corresponding area; at the same time, it maintains the independence of each area while enhancing the overall structural coherence and detail restoration capabilities.

[0158]

[0159] Where σ(·) represents the Sigmoid activation used to generate the smooth guiding weights. Indicates the convolution operation;

[0160] All the local refinement results are stitched together into a complete feature map:

[0161]

[0162] Subsequently, with attention enhancement features To merge:

[0163]

[0164] Finally, the output enhancement module generates the final output:

[0165]

[0166] in, Indicates the convolution operation; This indicates element-wise addition; Represents the concatenation operation of features;

[0167] This design explicitly models local structures under global guidance, effectively improving boundary integrity and structural consistency.

[0168] Example 2:

[0169] This embodiment proposes a language-guided structure-aware network architecture for camouflaged target detection based on the above embodiment 1, including a PVT-v2 backbone network, a CLIP model, a Fourier edge enhancement module (FEEM), a structure-aware attention module (SAAM), and a coarse-guided local refinement module (CGLRM).

[0170] The text prompts and RGB images are input into the CLIP model to obtain the text and visual features of the disguised target. The text-guided attention is used to generate a target mask, which guides the multi-scale features output by the PVT-v2 backbone network to focus on the potential target region.

[0171] The Fourier Edge Enhancement Module (FEEM) fuses multi-scale semantic features and extracts high-frequency information in the frequency domain to enhance edges. The resulting edge guidance features provide boundary cues for subsequent modules.

[0172] Furthermore, a Structure-Aware Attention Module (SAAM) is proposed to enhance the model's perception of target structures and boundaries;

[0173] Considering that relying solely on this module is insufficient to guarantee the overall coherence and local consistency of the objective, a coarse-guided local refinement module (CGLRM) is further proposed, employing a dual-branch structure:

[0174] One approach introduces channel attention and spatial attention to perform global perception modeling and generate attention-enhanced features;

[0175] The other branch divides the input features into four local regions;

[0176] Local structural refinement is performed under the guidance of attention-enhanced features, thereby effectively improving the boundary integrity and structural consistency of the camouflaged target segmentation results.

[0177] Based on the two embodiments above, the following specific experimental design is proposed:

[0178] Experimental setup

[0179] The method of this invention is evaluated on three publicly available benchmark datasets: CAMO, COD10K, and NC4K. These datasets provide unique conditions for training and evaluating COD models. The CAMO dataset contains 1,250 images covering 8 categories, focusing on camouflaged targets that blend well with natural backgrounds. COD10K expands the scale to 10,000 images across 78 categories. The NC4K dataset, containing 4,121 images, is designed to support target localization and ranking tasks. This invention uses the training sets of CAMO and COD10K for training and the test sets of these two datasets, as well as NC4K, for evaluation. Overall, these datasets provide a solid foundation for advancing COD research and model evaluation.

[0180] For quantitative evaluation, this invention selects four widely used indicators to assess the method of this invention: mean absolute error (MAE, M), weighted F-measurement, and weighted average error (F-measurement). Structural similarity measurement and average E measurement .

[0181] Experimental details

[0182] This invention is implemented using PyTorch, and all training and testing were performed on a single NVIDIA GeForce RTX 5000 GPU. During the training phase, the input images were uniformly scaled to 512×512, the Adam optimizer was used, training was performed for 25 epochs, and the batch size was 4; the initial learning rate was set to 1× The learning rate is dynamically adjusted using a poly decay strategy (decay power of 0.9). During the testing phase, the input is scaled to 512×512, and the output is restored to its original resolution after inference to complete the performance evaluation.

[0183] Comparative experiment

[0184] To fully validate the effectiveness of the proposed LGSANet, this invention systematically compared it with representative recent methods—BGNet, DINet, VSCoD, FSEL, CamoFormer, ESNet, SENet, UGDNett, BDCLNet, and CGCOD—on three mainstream COD benchmark datasets (CAMO, COD10K, and NC4K). See Table 1 for details.

[0185] Table 1 Quantitative Comparison

[0186] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A language-guided structure-aware method for camouflaged target detection, characterized by: At least the following steps are included: The visual backbone is a PVT-v2 network, using PVT-v2-b3 as the image feature extractor to extract features from the input image I∈R. H×W×3 The four-scale feature map is obtained, denoted as... ; To enhance the semantic information of the camouflaged target, a frozen CLIP model is further introduced. The text encoder of the CLIP model extracts text features from the text prompts, while the visual encoder extracts multi-scale visual features from layers 8, 16, and 24. Cross-scale alignment and fusion are performed using a feature pyramid network to obtain CLIP visual features Attn. CLIP visual features Attn and text features are then input into text-guided attention to generate a target mask. ; mask Used via MGFA operation And Attn's explicit feature enhancement, producing and This allows the model to focus on potential camouflage areas; Enhanced features First, input the Transformer to obtain the attn features; The Fourier edge enhancement module uses Using attn as input, generate edge-guided features E; In the next stage, the highest-level features will be... The attn features are concatenated and input into the coarse-guided local refinement module to obtain coarse-grained structural features. The coarse-grained structural features, along with E and attn features, are input into the structure-aware attention module to obtain structural enhancement features. Then, it is further refined through a local refinement module guided by coarse-grained principles, generating... ; For the remaining scales i=3,2,1, the refinement process is as follows: First, output the previous scale. Upsampling and comparing with the corresponding backbone features cascade; Then compressed into convolutional blocks ; Will Edge-guided features E and The features are fed into the structure-aware attention module, and then the output features of the structure-aware attention module are locally refined through a coarse-guided local refinement module to obtain the output. ; During the prediction phase, the network... Convolution outputs segmentation results at multiple scales, by and Intermediate prediction plots were obtained respectively. and And upsampled to the original resolution by interpolation; semantic-guided prediction generated by attn Edge prediction is obtained from E. This also includes a mask. ; The final output set is: O={ , , , , , } 2. The language-guided structure-aware method for camouflaged target detection according to claim 1, characterized in that: The operation of the Fourier edge enhancement module includes at least the following steps: Based on features and attn as input; First, the highest layer features Concatenate with attn and obtain high-level fused features through channel compression convolution: Subsequently, this feature is upsampled stepwise and concatenated with lower-level features, then fused via convolution, and the boundary representation is enhanced using the EdgeEnhancer module. in, This represents the concatenation operation of features. Indicates feature upsampling, This refers to the EdgeEnhancer module. Represents convolution; The EdgeEnhancer module is based on the idea of ​​average pooling difference: First, smooth the input features, then calculate their difference to extract high gradient regions; An edge weight map is then generated and fused with the original features using a residual approach, thereby explicitly enhancing boundary contrast in the spatial domain. Its mathematical form is as follows: Where Pool represents a 1×1 average pooling operation, and σ represents the Sigmoid function. Indicates batch normalization, Represents a 1×1 convolution. Represents the edge weight graph; To further extract high-frequency detail information of the target boundary, in the fused features Applying Fourier transform to explicitly capture the high-frequency response: in represent function Represents the Fourier transform; Finally, output edge-guided features: .

3. The language-guided structure-aware method for camouflaged target detection according to claim 1, characterized in that: The application of the structure-aware attention module includes at least the following steps: Given input features x ,Right now ; First, the Q, K, and V features are extracted from feature x through convolution; Then feature M is applied to Q, and edge-guided feature E is applied to K; in, Represents convolution; This indicates element-wise multiplication; Then Q, K, Transform the token representation from [B,C,H,W] into [B,HW,C], and then proceed. Vector normalization; To further reduce memory usage, approximate attention computation is used: First, perform Softmax on the transposed K, then... Multiply, and finally sum Multiply; in, This represents batch matrix multiplication; Finally, the output is reshaped to recover [B,C,H,W] and fused with the original input features via residual connections.

4. The language-guided structure-aware method for camouflaged target detection according to claim 1, characterized in that: The application of the coarse-guided local refinement module includes at least the following steps: Given an input feature map x∈ This invention first applies channel attention (CA) and spatial attention (SA) to extract global coarse guiding features: in, This indicates element-wise multiplication; Attention-enhancing feature G provides coarse-grained spatial guidance for the target; Next, the input feature x is spatially divided into four non-overlapping sub-regions: The attention-enhancing feature G was divided in the same way as This is to ensure that each local area is equipped with the corresponding global guidance information; Where σ(·) represents the Sigmoid activation used to generate the smooth guiding weights. Indicates the convolution operation; All the local refinement results are stitched together into a complete feature map: Subsequently, with attention enhancement features To merge: Finally, the output enhancement module generates the final output: in, Indicates the convolution operation; This indicates element-wise addition; This represents the splicing operation of features.

5. A language-guided structure-aware network architecture for camouflaged target detection, used in the language-guided structure-aware method for camouflaged target detection as described in any one of claims 1-4, characterized in that: It includes the PVT-v2 backbone network, CLIP model, Fourier edge enhancement module, structure-aware attention module, and coarse-guided local refinement module; The text prompts and RGB images are input into the CLIP model to obtain the text and visual features of the disguised target. The text-guided attention is used to generate a target mask, which guides the multi-scale features output by the PVT-v2 backbone network to focus on the potential target region. The Fourier edge enhancement module fuses multi-scale semantic features and extracts high-frequency information in the frequency domain to enhance the edges. The resulting edge guidance features provide boundary cues for subsequent modules. Furthermore, a structure-aware attention module is proposed to improve the model's perception of target structures and boundaries; Furthermore, a coarse-guided local refinement module is proposed, employing a dual-branch structure: One approach introduces channel attention and spatial attention to perform global perception modeling and generate attention-enhanced features; The other branch divides the input features into four local regions; Local structural refinement is performed under the guidance of attention-enhancing features.

Images