Camouflage target detection system based on attribute classification guided Transform model

Abstract

The invention discloses a camouflage target detection system based on an attribute classification guidance Transform model. The camouflage target detection system comprises a visual feature extraction module, an attribute classification guidance branch, a target segmentation branch and a multi-task joint loss function module. An attribute classification task is introduced into a Transform network, and modeling and guiding of semantic information of camouflage targets such as multiple targets, shielding and complex shapes are achieved; meanwhile, a detail texture extractor and a cross-layer attention mechanism are designed to enhance fine-grained structure perception, and a low-confidence region is optimized and adjusted through a feature self-adaptive refining strategy; and optimizing an attribute classification result and a camouflage target segmentation result through a multi-task joint loss function. According to the method, the problems of fuzzy segmentation boundaries and detail missing are effectively improved, and the discrimination and integrity of target features are enhanced; the semantic perception capability of the system is obviously improved, and the camouflage target can be identified more accurately.

Description

Technical Field

[0001] This invention relates to the fields of image processing and computer vision, and in particular to a camouflaged target detection system based on an attribute classification-guided Transformer model. Background Technology

[0002] Camouflaged object detection (COD) aims to accurately segment inconspicuous targets from camouflaged backgrounds. In classic COD frameworks, models typically rely on low-level visual features such as color and texture to segment targets. While these methods can be effective against simple backgrounds, their performance often degrades significantly in complex scenes due to blurred boundaries, target occlusion, and scale differences.

[0003] With the development of deep neural networks, existing COD methods can be broadly divided into two categories: methods based on convolutional neural networks (CNNs) and methods based on Transformers. CNN methods enhance saliency through hierarchical feature fusion, which has certain advantages in boundary detail modeling, but its global semantic understanding ability is limited; while Transformer methods improve global dependency modeling by utilizing attention mechanisms, making them more robust in complex backgrounds, but often lacking in detail representation.

[0004] However, both types of methods share common limitations: on the one hand, they generally ignore the semantic attributes of the camouflaged targets, such as the number of targets, occlusion status, and shape complexity, which can serve as effective prior information to assist in detection and segmentation; on the other hand, existing methods mostly impose constraints only on the final prediction results, lacking direct optimization of intermediate feature layers, making it difficult to improve the overall segmentation accuracy through feature-prediction collaboration.

[0005] Therefore, a COD framework that can explicitly model semantic attributes and simultaneously optimize features and predictions is needed to overcome the performance bottleneck of existing methods in complex camouflage scenarios. Summary of the Invention

[0006] Purpose of the invention: The present invention aims to provide a camouflaged target detection system based on attribute classification-guided Transformer model to improve the detection and segmentation accuracy of camouflaged targets, especially in complex scenarios such as multiple targets, small targets, occlusion, and unclear boundaries.

[0007] Technical solution: The camouflaged target detection system based on attribute classification-guided Transformer model described in this invention includes:

[0008] The visual feature extraction module is used to receive the input image and extract multi-layer feature maps and corresponding category labels;

[0009] The attribute classification guidance branch module ACGB is used to classify the attributes of possible camouflaged targets in the image using the category labels.

[0010] The Target Segmentation Branch Module (OSB) includes: a Detail Texture Extractor (DTE), a Cross-Agent Attention Module (CAA), and a Feature Adaptive Refinement Module (FAR). The DTE extracts detailed image information; the CAA fuses multi-level features to enhance the salience of camouflaged targets; and the FAR guides low-confidence regions through high-confidence regions for feature optimization.

[0011] The multi-task joint loss function module is used to jointly optimize the attribute classification result and the camouflaged target segmentation result.

[0012] Furthermore, the attributes include multi-target (MO), large target (BO), small target (SO), field of view extension (OV), occlusion (OC), complex shape (SC), and unclear boundary (IB).

[0013] Furthermore, the visual feature extraction module is a visual encoder based on the Transformer architecture.

[0014] Furthermore, the attribute classification guidance branch ACGB uses a multi-layer linear layer and the ReLU activation function to reduce the dimensionality of the category labels and output the attribute classification results.

[0015] Furthermore, the detail texture extractor (DTE) includes a receptive field module (RFB), two layers of deformable convolutional layers (DCNv4), and two layers of dynamic snake-like convolutional layers (SnakeConv); the input feature map F 0 ∈R B×C×H×W Enhance receptive field and multi-scale contextual information through RFB:

[0016] F 1 =RFB(F 0 )

[0017] For F 1 Perform the first layer of DCNv4 deformable convolution and add it to the input residual to achieve feature fusion:

[0018] F 2 =DCNv4(F 1 )+F 1

[0019] F 2 Two dynamic serpentine convolutions in two directions are input to extract texture and edge information in the horizontal and vertical directions respectively, and then concatenated along the channel dimension:

[0020] F 3=Cat(SnakeConv xy (F 2 ),SnakeConv yx (F 2 ))

[0021] For the spliced ​​feature F 3 Perform 3×3 convolution and normalized BN processing:

[0022] F 4 =ConvBN(F 3 )

[0023] DCNv4 is used again to enhance the feature representation, and it is combined with F 4 The residues are added together to form residual connections, and finally, ReLU activation is applied to obtain the output feature map.

[0024] F = ReLU(DCNv4(F) 4 )+F 4 )

[0025] Where B is the batch size, C is the number of channels, and H and W are the spatial dimensions; the shape of the output F is the same as that of the input F. 0 same.

[0026] Furthermore, the implementation process of the cross-proxy attention module (CAA) is as follows:

[0027] The two adjacent feature maps are F i and F i+1 F i+1 For higher-level features with stronger semantics, F i For lower-level features with higher spatial resolution; the computational process of CAA is formalized as follows:

[0028] CAA(F i ,F i+1 )=AA(Q,K,V,A)

[0029] The query is defined as Q = F. i +W q F i+1 The key and value are K = F. i +W k F i+1 V = F i +W v F i+1 The proxy vector is A = AF i +W q F i+1 W q W k W v It is a learnable linear mapping matrix; This represents a proxy token generated from low-level features.

[0030] Furthermore, the feature adaptive refinement module (FAR) is implemented as follows:

[0031] Obtain the final prediction graph P∈R H×W With the corresponding true mask G∈{0,1} H×W And define the pseudo-label as TF = G·(P>0), where · represents element-wise multiplication;

[0032] For each pixel i, calculate its binary cross-entropy loss l i =L bce (p i ,tf i And press l i Sort the pixels from smallest to largest;

[0033] The corresponding shallow feature map F∈R N×C Where N = H × W, rearranged into the sorted characteristic sequence F seq ;

[0034] Based on the ranking results, the low-confidence region feature F is selected. bad =F seq [αN:βN] and high-confidence region features F well =F seq [γN:δN]; where α=1 / 8, β=1 / 2, γ=3 / 4, δ=1;

[0035] Calculate the high-confidence feature mean The target foreground alignment loss is obtained by minimizing the low-confidence region and its mean square error.

[0036] L obj =MSE(F bad ,μ well )

[0037] The target-background alignment loss L is calculated in the same manner. bgc The feature alignment loss is obtained as: L far =L obj +L bgc .

[0038] Furthermore, the multi-task joint loss function includes segmentation loss, masquerading attribute weighted classification loss, and feature alignment loss:

[0039] Attribute-weighted classification loss: L attr =L bce (P attr G attr )

[0040] Segmentation loss: L seg (P,G)=L bce (P,G)+L iou (P,G), where P is the segmentation prediction map and G is the binary ground truth mask; L bce It is a binary cross-entropy loss function, L iou It is the intersection-union ratio loss function;

[0041] Feature alignment loss: L far =L obj +L bgc Among them, L obj For the target foreground alignment loss, L bgc Alignment loss with target background;

[0042] The final loss function is:

[0043]

[0044] Among them, P i It is the predicted i-th layer segmentation map, i∈{1,2,3,4}.

[0045] Furthermore, the camouflage attribute weighted classification loss is defined as:

[0046]

[0047] Wherein, weight w i Normalization is performed based on the reciprocal of the frequency of attribute i in the training set, i.e.:

[0048]

[0049] in, n i Let N be the number of samples for attribute i, and N be the total number of samples.

[0050] Furthermore, the system also includes an input and data processing module, which unifies the size of feature maps at all levels before multi-layer feature fusion and output prediction, thereby improving spatial alignment and semantic consistency.

[0051] Beneficial Effects: Compared with existing technologies, the beneficial effects of this invention are as follows: By utilizing target attributes to guide the segmentation process, this invention significantly improves the semantic perception capability of the system, enabling more accurate identification of camouflaged targets; the introduction of a feature alignment mechanism effectively improves the problems of blurred segmentation boundaries and missing details, enhancing the discriminativeness and completeness of target features; experimental results on multiple standard datasets (such as COD10K, CAMO, CHAMELEON, NC4K) show that this invention significantly outperforms existing mainstream methods, exhibiting higher segmentation accuracy and robustness; this invention has good scalability and generalization ability, and can be widely applied to various scenarios such as medical image analysis, video surveillance, and natural image processing. Attached Figure Description

[0052] Figure 1 This is a schematic diagram of the framework of the present invention;

[0053] Figure 2 This is a schematic diagram of the structure of the detail texture extractor (DTE).

[0054] Figure 3 The flowchart for the Feature Adaptive Refinement (FAR) strategy is shown below.

[0055] Figure 4 This is a schematic diagram illustrating the segmentation of camouflaged targets with different attributes according to the present invention; wherein, (a) segmentation results under the small target (SO) attribute; (b) segmentation results under the multi-target (MO) scenario; (c) segmentation results under the complex shape (SC) scenario; (d) segmentation results under the field of view extension (OV) scenario; and (e) segmentation results under the occlusion (OC) scenario.

[0056] Figure 5 This diagram illustrates a comparison between the segmentation results of the present invention and existing methods for camouflaged targets with different attributes. Detailed Implementation

[0057] The present invention will now be described in further detail with reference to the accompanying drawings.

[0058] like Figure 1As shown, this invention proposes a camouflaged target detection system based on an attribute classification-guided Transformer model, including a visual feature extraction module, preferably Vision Transformer (ViT) or DINOv2, which extracts multi-layer feature maps and class tokens; an attribute classification-guided branch (ACGB) performs multi-label attribute classification on the class tokens; an object segmentation branch (OSB) includes a detail texture extractor (DTE), a cross-agent attention module (CAA), and a feature adaptive refinement module (FAR), outputting multi-level segmentation maps, and jointly trained with the attribute branch; the system also includes an input and data processing module and a multi-task joint loss function module; the specific implementation of each module is as follows:

[0059] (1) Input and Data Processing:

[0060] The input RGB image (arbitrary resolution) is uniformly preprocessed to 448×448 (width×height) during both training and inference phases; the RGB image is standardized according to ImageNet mean and variance (mean=[0.485,0.456,0.406], std=[0.229,0.224,0.225]). The original binary ground truth mask ({0,255}) is normalized to [0,1] and scaled and resampled to 448×448. Furthermore, the mask is downsampled to 32×32 for multi-level segmentation supervision or as a reference for FAR.

[0061] (2) Visual feature extraction module:

[0062] Using DINOv2 Small / Base as the visual feature extraction module, taking DINOv2 Small as an example, the input image size is 448×448, the patch size is 14, and the final 4 layers of class tokens and patch tokens are obtained, with a dimension of D=384. The patch tokens are then reshaped into the shape of the feature map through a reshape operation to obtain the feature map.

[0063] (3) Attribute Category Bootstrap Branch (ACGB):

[0064] First, the class tokens extracted by DINOv2 are concatenated together, then passed through two linear layers and a ReLU function, and finally through a linear layer to output the logits corresponding to the seven attributes; finally, the Sigmoid activation function is used to convert them into probabilities.

[0065] The seven attributes include: Multiple Objects (MO), Big Object (BO), Small Object (SO), Out-of-View (OV), Occlusion (OC), Shape Complexity (SC), and Indistinct Boundary (IB).

[0066] A dummy attribute weighted classification loss is used to constrain the attribute classification and guide the branch to make correct attribute probability vector predictions, thereby optimizing the DINOv2 feature representation:

[0067]

[0068] Wherein, weight w i It is calculated by normalizing the inverse of the frequency of attribute i in the training set, that is:

[0069]

[0070] in, n i Let N be the number of samples for attribute i, and N be the total number of samples.

[0071] (4) Target Segmentation Branch (OSB): includes Detail Texture Extractor (DTE), Cross-Agent Attention Module (CAA), and Feature Adaptive Refinement Module (FAR).

[0072] like Figure 2 As shown, the Detail and Texture Extractor (DTE) proposed in this invention is used to enhance the texture and boundary representation of camouflaged targets, and is particularly suitable for complex multi-target scenes. This module includes a Receptive Field Block (RFB), two layers of deformable convolutional DCNv4, two layers of dynamic snake convolutional (SnakeConv), a 3×3 convolution, batch normalization (BN), and a ReLU activation function, with the following structure:

[0073] Input feature map F 0 ∈R B×C×H×W Where B is the batch size, C is the number of channels, and H and W are the spatial dimensions. First, the receptive field and multi-scale contextual information are enhanced using the RFB module:

[0074] F 1 =RFB(F 0 )

[0075] For F 1Perform the first layer of DCNv4 deformable convolution and add it to the input residual to achieve feature fusion:

[0076] F 2 =DCNv4(F 1 )+F 1

[0077] This step helps capture global semantic information while preserving the low-frequency structure of the original features.

[0078] F 2 Two dynamic snake-like convolutions (X-SnakeConv and Y-SnakeConv) in two directions are input to extract texture and edge information in the horizontal and vertical directions, respectively, and then concatenated along the channel dimension:

[0079] F 3 =Cat(SnakeConv xy (F 2 ),SnakeConv yx (F 2 ))

[0080] For the spliced ​​feature F 3 Perform 3×3 convolution and normalized BN processing:

[0081] F 4 =ConvBN(F 3 )

[0082] DCNv4 is used again to enhance the feature representation, and it is combined with F 4 The residues are added together to form residual connections, and finally, ReLU activation is applied to obtain the output feature map.

[0083] F = ReLU(DCNv4(F) 4 )+F 4 )

[0084] Wherein, the shape of the output F is the same as that of the input F 0 Similarly, it can be directly used as input to downstream cross-agent attention modules (CAA) or split headers.

[0085] In the implementation process, the RFB module is used to expand the receptive field and fuse multi-scale information, which can be achieved by combining standard convolution and variable receptive field convolution; DCNv4 uses convolution kernels with learnable offsets, which can dynamically adjust the receptive region to capture global semantic information; SnakeConv in the X and Y directions performs directional dynamic convolution on features to enhance boundary and texture details; all convolution operations use 3×3 convolution kernels, and appropriate padding is used to keep the spatial size unchanged to ensure consistent spatial resolution of feature maps.

[0086] This invention introduces a Cross Agent Attention (CAA) module during feature fusion to achieve efficient information interaction and guidance between multi-layer features. Specifically, high-level features contain rich semantic information, while low-level features retain more details and boundary information; both are equally important for camouflaged target detection. To fully utilize these two types of features, CAA borrows the idea of ​​Agent Attention, using high-level features as agents to guide the updating of low-level features.

[0087] In the implementation process, let F be the feature maps of two adjacent layers. i and F i+1 Among them, F i+1 For higher-level features with stronger semantics, F i For lower-level features with higher spatial resolution, the CAA calculation process can be formalized as follows:

[0088] CAA(F i ,F i+1 )=AA(Q,K,V,A)

[0089] The query is defined as Q = F. i +W q F i+1 The key and value are K = F. i +W k F i+1 V = F i +W v F i+1 The proxy vector is Here, W q W k W v It is a learnable linear mapping matrix; This represents a proxy token generated from low-level features.

[0090] Through the above construction, high-level features play a biasing guiding role in attention calculation, enabling low-level features to better align with high-level semantics while preserving details, thereby improving the salience of camouflaged targets. Unlike traditional feature fusion methods based on downsampling, CAA retains the entire token set during calculation, thus exhibiting higher fidelity during information interaction. To balance computational efficiency, this invention introduces the CAA module only between the highest three feature layers of the visual feature extraction module, achieving cross-layer semantic-detail interaction with minimal computational overhead, thereby enhancing the overall performance of camouflaged target segmentation.

[0091] The prediction head is a convolutional layer that reduces the dimension of the feature map to 1 and outputs a prediction map. Finally, upsampling is applied to adjust the prediction map to a size of 448×448 to facilitate subsequent calculation of the segmentation loss.

[0092] like Figure 3 As shown, the Feature Adaptive Refinement (FAR) module guides feature optimization in low-confidence regions by using high-confidence regions. The FAR module sorts the predicted loss values, adaptively filters pixel regions with low confidence, and combines this with the mean feature value of high-confidence regions for guided optimization, thereby improving the discriminative power of the target features. The feature optimization process performed by the FAR module is as follows:

[0093] Obtain the final prediction graph P∈R H×W With the corresponding true mask G∈{0,1} H×W The pseudo-label is defined as TF = G·(P>0), where · represents element-wise multiplication. For each pixel i, its binary cross-entropy loss l is calculated. i =L bce (p i ,tf i And press l i Sort the pixels from smallest to largest. Then, sort the corresponding shallow feature maps F∈R. N×C (where N = H × W) is rearranged into the sorted characteristic sequence F. srq Based on the ranking results, the low-confidence region feature F is selected. bad =F seq [αN:βN] and high-confidence region features F well =F seq [γN:δN]; where α = 1 / 8, β = 1 / 2, γ = 3 / 4, δ = 1. Calculate the mean of the high-confidence features. The target foreground alignment loss is obtained by minimizing the low-confidence region and its mean square error.

[0094] L obj =MSE(F bad ,μ well )

[0095] Calculate L in the foreground and background regions in the same manner. obj With L bgc The final combination yields the total feature alignment loss:

[0096] L far =L obj +L bgc .

[0097] (5) Multi-task joint loss (training objective):

[0098] The system optimization employs a multi-task joint loss function to simultaneously constrain the accuracy of camouflaged target segmentation, attribute classification accuracy, and feature consistency. The joint loss function comprises three parts: segmentation loss, camouflage attribute weighted classification loss, and feature alignment loss. Specifically:

[0099] Attribute-based weighted classification loss L attr =L bce (P attr G attr Feature alignment loss L far =L obj +L bgc The segmentation loss is used to optimize the target segmentation accuracy of the model. For each layer of segmentation prediction map P and the corresponding binary ground truth mask G, a binary cross-entropy loss L is calculated. bce (P,G) and crossover ratio loss L iou The sum of (P,G) is used to simultaneously consider pixel-level prediction accuracy and region overlap consistency.

[0100] L seg (P,G)=L bce (P,G)+L iou (P,G)

[0101] The final multi-task joint loss function is the sum of the above three terms:

[0102]

[0103] Among them, P i G is the predicted segmentation map of the i-th layer (i∈{1,2,3,4}), and G is the binary ground truth mask.

[0104] The system proposed in this invention uses the COD10K and CAMO training sets as training data. During the training phase, a batch size of 36 is employed, and data augmentation operations, including horizontal flipping, rotation, and transposition, are applied to the input images to improve the model's generalization ability. All training and testing images are uniformly resized to 448×448 to ensure consistent input scale. The network is optimized using the Adam optimizer with an initial learning rate of 0.0001 and a total training duration of 100 epochs. The learning rate is decayed by 10% every 40 epochs to achieve stable convergence. During the testing phase, the input images are also resized to 448×448, but no further data augmentation is performed. After the model training is complete, it can be deployed in an automatic camouflage target detection system for inference applications.

[0105] During the inference phase, the input image is first resized to 448×448 and fed into the visual feature extraction module to obtain the final four layers of patch tokens and class tokens. The patch tokens are then reshaped into spatial feature maps for subsequent processing. Next, the attribute classification guidance branch (ACGB) uses the class tokens to predict attribute probability vectors, predicting the attribute probabilities of camouflaged targets in the image. Simultaneously, the object segmentation branch (OSB) outputs a multi-scale segmentation probability set P based on the patch tokens. i (where i∈{1,2,3,4}), capturing spatial information of the target at different scales. To obtain the final segmentation result, the highest resolution prediction P4 is selected as the output segmentation map, which can be directly used for camouflaged target detection tasks. That is, the output segmentation map can be used as a mask for target localization or further analysis.

[0106] Figure 4 This paper presents schematic diagrams illustrating the segmentation results of this invention for camouflaged target scenes with different attributes, including small targets (SO), multiple targets (MO), shape complexity (SC), field of view extension (OV), and occlusion (OC). By presenting the original image containing camouflaged targets and the corresponding segmentation results, the model's ability to segment targets in various camouflaged scenes can be intuitively observed, helping to verify the effectiveness of the method in camouflaged target segmentation tasks.

[0107] To verify the effectiveness of the present invention, Figure 5 This paper compares the present invention with two state-of-the-art models in the field of camouflaged target segmentation: FSPNet and FEDER. From... Figure 5 The results of different algorithms on the target region can be seen intuitively. When processing diverse and partially complex scene images, such as those containing faces and marine life, FSPNet and FEDER suffer from inaccurate masking, misidentification, or omission of targets (e.g., FEDER misidentifies faces, and some model mask shapes deviate from the real targets in marine life images). In contrast, the mask obtained by this invention more closely matches the actual target contour, resulting in higher positioning accuracy. This invention exhibits higher robustness and reliability in scenes containing different types (SO, SC, MO, etc.), enabling more accurate target masking and localization, and distinguishing targets from the background and other interfering elements.

[0108] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A camouflaged target detection system based on attribute classification-guided Transformer model, characterized in that, include: The visual feature extraction module is used to receive the input image and extract multi-layer feature maps and corresponding category labels; The attribute classification guidance branch module ACGB is used to classify the attributes of possible camouflaged targets in the image using the category labels. The Target Segmentation Branch Module (OSB) includes: a Detail Texture Extractor (DTE), a Cross-Agent Attention Module (CAA), and a Feature Adaptive Refinement Module (FAR). The DTE extracts detailed image information; the CAA fuses multi-level features to enhance the salience of camouflaged targets; and the FAR guides low-confidence regions through high-confidence regions for feature optimization. The multi-task joint loss function module is used to jointly optimize the attribute classification result and the camouflaged target segmentation result.

2. The camouflaged target detection system based on attribute classification-guided Transformer model according to claim 1, characterized in that, The aforementioned attributes include multi-target (MO), large target (BO), small target (SO), field of view extension (OV), occlusion (OC), complex shape (SC), and unclear boundary (IB).

3. The camouflaged target detection system based on attribute classification-guided Transformer model according to claim 1, characterized in that, The visual feature extraction module is a visual encoder based on the Transformer architecture.

4. The camouflaged target detection system based on attribute classification-guided Transformer model according to claim 1, characterized in that, The attribute classification guidance branch ACGB uses a multi-layer linear layer and ReLU activation function to reduce the dimensionality of the category labels and output the attribute classification results.

5. The camouflaged target detection system based on attribute classification-guided Transformer model according to claim 1, characterized in that, The detail texture extractor (DTE) includes a receptive field module (RFB), two layers of deformable convolutional layers (DCNv4), and two layers of dynamic snake-like convolutional layers (SnakeConv); the input feature map is F. 0 ∈R B×C×H×W Enhance receptive field and multi-scale contextual information through RFB: F 1 =RFB(F 0 ) For F 1 Perform the first layer of DCNv4 deformable convolution and add it to the input residual to achieve feature fusion: F 2 =DCNv4(F 1 )+F 1 F 2 Two dynamic serpentine convolutions in two directions are input to extract texture and edge information in the horizontal and vertical directions respectively, and then concatenated along the channel dimension: F 3 =Cat(SnakeConv xy (F 2 ),SnakeConv yx (F 2 )) For the spliced ​​feature F 3 Perform 3×3 convolution and normalized BN processing: F 4 =ConvBN(F 3 ) DCNv4 is used again to enhance the feature representation, and it is combined with F 4 The residues are added together to form residual connections, and finally, ReLU activation is applied to obtain the output feature map. F=ReLU(DCNv4(F 4 )+F 4 ) Where B is the batch size, C is the number of channels, and H and W are the spatial dimensions; the shape of the output F is the same as that of the input F. 0 same.

6. The camouflaged target detection system based on attribute classification-guided Transformer model according to claim 1, characterized in that, The implementation process of the cross-proxy attention module (CAA) is as follows: The two adjacent feature maps are F i and F i+1 F i+1 For higher-level features with stronger semantics, F i For lower-level features with higher spatial resolution; The calculation process of CAA is formalized as follows: CAA(F i ,F i+1 )=AA(Q,K,V,A) The query is defined as Q = F. i +W q F i+1 The key and value are K = F. i +W k F i+1 V = F i +W v F i+1 The proxy vector is W q W k W v It is a learnable linear mapping matrix; This represents a proxy token generated from low-level features.

7. The camouflaged target detection system based on attribute classification-guided Transformer model according to claim 1, characterized in that, The Feature Adaptive Refinement (FAR) module is implemented as follows: Obtain the final prediction graph P∈R H×W With the corresponding true mask G∈{0,1} H×W And define the pseudo-label as TF = G·(P>0), where · represents element-wise multiplication; For each pixel i, calculate its binary cross-entropy loss l i =L bce (p i ,tf i And press l i Sort the pixels from smallest to largest; The corresponding shallow feature map F∈R N×C Where N = H × W, rearranged into the sorted characteristic sequence F seq ; Based on the ranking results, the low-confidence region feature F is selected. bad =F seq [αN:βN] and high-confidence region features F well =F seq [γN:δN]; where α=1 / 8, β=1 / 2, γ=3 / 4, δ=1; Calculate the high-confidence feature mean The target foreground alignment loss is obtained by minimizing the low-confidence region and its mean square error. L obj =MSE(F bad ,μ well ) The target-background alignment loss L is calculated in the same manner. bgc The feature alignment loss is obtained as: L far =L obj +L bgc .

8. The camouflaged target detection system based on attribute classification-guided Transformer model according to claim 1, characterized in that, The multi-task joint loss function includes segmentation loss, masquerading attribute weighted classification loss, and feature alignment loss: Attribute-weighted classification loss: L attr =L bce (P attr G attr ) Segmentation loss: L seg (P,G)=L bce (P,G)+L iou (P,G), where P is the segmentation prediction map and G is the binary ground truth mask; L bce It is a binary cross-entropy loss function, L iou It is the intersection-union ratio loss function; Feature alignment loss: L far =L obj +L bgc Among them, L obj For the target foreground alignment loss, L bgc Alignment loss with target background; The final loss function is: Among them, P i It is the predicted i-th layer segmentation map, i∈{1,2,3,4}.

9. The camouflaged target detection system based on attribute classification-guided Transformer model according to claim 8, characterized in that, The camouflage attribute weighted classification loss is defined as: Wherein, weight w i Normalization is performed based on the reciprocal of the frequency of attribute i in the training set, i.e.: in, n i Let N be the number of samples for attribute i, and N be the total number of samples.

10. The camouflaged target detection system based on attribute classification-guided Transformer model according to claim 1, characterized in that, The system also includes an input and data processing module, which unifies the size of feature maps at all levels before multi-layer feature fusion and output prediction, thereby improving spatial alignment and semantic consistency.

Images