Camouflage target detection method based on edge information and coarse segmentation

Abstract

The invention discloses a camouflage target detection method based on edge information and coarse segmentation, and the method comprises the steps: extracting initial features from each layer of an image through a backbone network, and recording the initial features as f1-f5; the features f2-f5 are input into an edge sensing module, edge labels guide to effectively extract edge information of the camouflage target from each layer, and an edge feature map is generated. The features f1-f5 are input into a feature enhancement module, and the preliminary features extracted by the backbone network are further refined. The rough segmentation module takes f1-f5 as input, the deep features are gradually fused into the shallow features through the feature enhancement module, and the fused features are processed through convolution to generate a rough segmentation image; and the semantic integration module integrates useful information extracted from the edge feature map and the rough segmentation map into output features of the feature enhancement module, and finally a prediction map is obtained. According to the method, the segmentation precision and detail performance can be remarkably improved, and the camouflage target can be more accurately identified and segmented in a complex background.

Description

Technical Field

[0001] The present invention relates to target detection, and in particular to a camouflaged target detection method based on edge information and coarse segmentation. The present invention can be applied to camouflaged target detection in autonomous driving, intelligent security, and military reconnaissance to improve the recognition accuracy and response efficiency of concealed or camouflaged objects, and belongs to the field of intelligent recognition technology. Background Art

[0002] Camouflage is a strategy used by organisms in nature to hide themselves by blending in with their surroundings. Due to their high similarity to the background, it is usually difficult for predators and prey to detect them. The task of camouflaged object detection (COD) is to segment out targets that are difficult for the human eye to identify, especially those that blend in closely with the complex background. COD technology has been widely used in many fields, including medical image segmentation (such as polyp detection, lung infection diagnosis, and retinal image segmentation), industrial defect detection, species discovery and identification, urban planning (such as road recognition), image retrieval, military (camouflage pattern design), and agriculture (such as pest identification), and has therefore attracted more and more attention.

[0003] Early COD methods mainly rely on handcrafted low-level features, which are easily disturbed by complex backgrounds, resulting in low detection accuracy. For example, traditional image processing techniques may fail to effectively separate the target from the background because the texture, color, and shape of the background are too similar to the target. To address this problem, many deep learning-based COD methods have gradually emerged. These methods use the powerful representation ability of neural networks to automatically learn more complex feature patterns from large amounts of data, showing great potential. Current COD methods perform well in detecting single camouflaged targets and can accurately identify camouflaged targets in relatively simple scenes. However, when faced with complex scenes such as multiple targets, occlusions, and overlapping objects, the performance of these methods drops significantly. Especially when the foreground and background are highly similar, even advanced algorithms such as TPRNet and C2FNet fail to accurately reproduce the complete structure of the camouflaged target when trying to detect the camouflaged target in the camouflaged area. This is mainly because these advanced methods fail to fully recognize the importance of using edge information in camouflaged target detection, which can provide key clues about the target outline and help the algorithm define the target boundary more accurately. In addition, advanced methods such as EAMNet have limited capabilities in dealing with camouflaged targets with slight occlusion, which leads to missed detections and false detections, and ultimately produces inaccurate prediction maps. This is mainly due to inaccurate edge prior guidance and insufficient position prediction, that is, the algorithm fails to effectively use the existing edge information to guide the localization and segmentation of the target. Therefore, an ideal camouflaged target detection method should use more representative features from multiple perspectives to compensate for the limitations of human vision.

[0004] Although existing camouflaged target detection technologies have achieved certain achievements, they still face significant challenges in several key areas. First, the high similarity between camouflaged targets and backgrounds leads to insufficient detection capabilities in complex backgrounds, especially in natural environments, where the diversity and complexity of the background further exacerbate this problem. Second, for partially occluded or overlapping camouflaged targets, the processing capabilities of existing technologies are limited, and detection failures or incomplete segmentation often occur. In addition, edge information plays a vital role in camouflaged target detection, but many existing methods do not fully explore or effectively utilize this information, resulting in unsatisfactory detection results. To address these issues, future research needs to focus on developing more efficient feature extraction and representation methods to improve the model's robustness to complex backgrounds and the accuracy of multi-target detection; at the same time, edge information and other auxiliary information should be fully utilized to improve the accuracy of target detection and segmentation. Finally, exploring more effective supervised learning methods and reducing training time will help promote the development and application of camouflaged target detection technology. Summary of the invention

[0005] In view of the problems of high training cost and narrow application scope in most existing camouflaged target detection methods, especially the segmentation difficulties caused by large shape differences between different camouflaged targets, blurred boundaries and small target segmentation objects, the purpose of the present invention is to provide a camouflaged target detection method based on edge information and coarse segmentation. The present invention can significantly improve the segmentation accuracy and detail performance, can more accurately identify and segment camouflaged targets in complex backgrounds and multi-target scenes, and the target recognition is fast and efficient.

[0006] The technical solution of the present invention is achieved in this way:

[0007] A camouflaged target detection method based on edge information and coarse segmentation, the steps are as follows:

[0008] 1) Training a disguised target detection model; the disguised target detection model includes a backbone network, an edge perception module, a feature enhancement module, a rough segmentation module and a semantic integration module;

[0009] 1.1) The original RGB image dataset is fed into a predefined backbone network; the backbone network consists of a series of convolutional layers that decompose the input image into five different resolution levels by gradually downsampling it; at each level, preliminary features of different levels are extracted from the image through the application of convolution operations and activation functions, denoted as f i (i=1, 2, 3, 4, 5);

[0010] 1.2) Input features f2~f5 into the edge perception module. The edge labels guide the effective extraction of edge information of the camouflaged target from each layer. The generated edge detection feature map is expressed as

[0011] At the same time, features f1~f5 are respectively input into the feature enhancement module, through which redundant noise is removed, the receptive field is expanded, and the number of channels is reduced to 64, and the initial features extracted by the backbone network are further refined. The output of the feature enhancement module is expressed as f′ i (i=1,2,3,4,5);

[0012] 1.3) The coarse segmentation module is used to generate a rough segmentation map of the camouflaged target; the coarse segmentation module takes the multi-level feature maps f1, f2, f3, f4 and f5 as input, and gradually fuses the deep features into the shallow features through the feature enhancement module, and processes the fused features through a 3×3 convolution layer to generate a coarse segmentation map, which is expressed as

[0013] 1.4) The edge detection feature map generated by the edge perception module The output f′ of the feature enhancement module i (i=2,3,4,5) and the coarse segmentation map generated by the coarse segmentation module The corresponding input is sent to the semantic integration module; the semantic integration module combines the high-level and low-level features of each layer to output the corresponding prediction graphs respectively, and then the prediction graphs output by the high-level semantic integration module are used as the inputs of the low-level semantic integration modules in turn. The output of the lowest-level semantic integration module is the final prediction graph, which is the output of the disguised target detection model;

[0014] 2) The RGB image to be detected is input into the disguised target detection model, and the output of the disguised target detection model is the disguised target image obtained by detection.

[0015] Furthermore, in step 1.2), the feature enhancement module consists of four sub-branches B k (k=1, 2, 3, 4) and a main branch B O Each sub-branch contains a depth-wise separable convolution, and the input of the feature enhancement module is f i (i=1, 2, 3, 4, 5), the output is expressed as:

[0016]

[0017] Furthermore, in step 1.2), the edge perception module is used to fuse high-level and low-level features to generate auxiliary edge information; in the edge perception module, a 1×1 convolution is used to adjust the number of channels to 64, and the features (f3 to f5) are upsampled to match the dimension of f2, and channel splicing is performed; then, features of adjacent scales are spliced ​​to retain multi-scale information.

[0018] Furthermore, the edge perception module includes multiple feature filtering blocks, which are used to filter the feature f m and the feature f of its adjacent scale n Splice and get f mn ; The concatenated feature f mn It is processed by two independent branches, one of which includes a 1×1 convolutional layer and a sigmoid activation function; considering the input tensor f mn =[f 1，1 , f 2 ，2 , ..., f i，j , ..., f H，W ] is a slice representation, where Represents the local features at the spatial position (i, j), i∈{1, 2, ..., H}, j∈{1, 2, ..., W}; performs spatial compression through convolution operations to generate a projection tensor q∈R H×W ; The projection tensor is adjusted to the range [0,1] through the Sigmoid layer σ(·) to generate the activation of the spatial region The other branch consists of a 1×1 convolutional layer followed by two 3×3 convolutional layers; each convolutional layer is followed by batch normalization and ReLU activation layers to generate the output feature map f′ mn ∈R H×W×2 ; Then apply the Softmax function to the feature map to obtain two weight vectors W m and W n , used to fuse feature maps, the expression is The final output f″ mn It is expressed as:

[0019] f″ mn =U+V#(2)

[0020] in, Represents element multiplication; Finally, the multi-scale feature maps are integrated through residual connections, and the expression is as follows:

[0021]

[0022] Furthermore, in step 1.3), the specific processing process of the rough segmentation module is described as follows:

[0023]

[0024] Further, in step 1.4), the semantic integration module will detect the feature map from the edge and the rough segmentation map The useful information extracted from the feature enhancement module is integrated into the output feature f′ i (i=2,3,4,5), the processing process is described as follows:

[0025]

[0026] Combined with the input of the high-level semantic integration module, the output of the low-level semantic integration module is expressed as follows:

[0027]

[0028] Furthermore, in step 1), when training the camouflaged target detection model, the distinction between the foreground and the background is improved by a loss function, wherein the loss function includes a prediction loss L p and marginal loss L e ; Prediction loss L p It is expressed as follows:

[0029]

[0030] in, is the weighted binary cross entropy loss; is the weighted intersection-over-union loss; G t The mask of the camouflaged target, P i (i=2, 3, 4, 5) represents four camouflaged target masks generated by the edge-aware module;

[0031] Edge loss L e Use Dice loss L dice To automatically handle the significant imbalance between positive and negative samples;

[0032] The total loss function is expressed as follows:

[0033] L total =L p +λL dice (P e , G e )#(7)

[0034] Among them, λ is a hyperparameter used to balance the relationship between object edge loss and overall loss; L dice is the Dice loss; G e represents the edge of the camouflaged target, P e Represents the prediction of the edge of the camouflaged target.

[0035] The present invention proposes a new fast and accurate camouflage segmentation network model, which effectively guides the network to deeply explore and utilize representative edge features to avoid the loss of edge details. At the same time, coarse segmentation makes it possible to quickly search within the range of the camouflaged target and determine the approximate position of the camouflaged target, so that subsequent fine processing can be concentrated on these areas to avoid blind searching in the entire image. Finally, the edge information is effectively combined with the coarse segmentation map, which can significantly improve the segmentation accuracy and detail performance. By utilizing the key contour features provided by the edge information to supplement the missing details in the coarse segmentation map, this network design alleviates the challenges associated with the loss of edge details and incomplete segmentation, thereby improving the detection accuracy.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] Compared with the existing camouflage target detection COD technology, the present invention significantly improves the accuracy and detail performance of segmentation by effectively combining edge information with a rough segmentation map, and can more accurately identify and segment camouflage targets in complex backgrounds and multi-target scenes. Secondly, this combination method speeds up the processing speed, reduces the consumption of computing resources, and makes the algorithm more efficient in real-time applications. In addition, the present invention performs well in processing objects with blurred boundaries, large shape differences, and small target segmentation objects, effectively solving the common problems of incomplete segmentation and false detection in the prior art, and enhancing the robustness and adaptability of the algorithm. In summary, the present invention not only achieves a breakthrough in technology, but also shows significant advantages in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 -Schematic diagram of the model architecture of the camouflaged target detection method of the present invention.

[0039] Figure 2 -Schematic diagram of the structure of the feature enhancement module of the present invention.

[0040] Figure 3 -Schematic diagram of the edge perception module structure of the present invention.

[0041] Figure 4 -Schematic diagram of the rough segmentation module structure of the present invention.

[0042] Figure 5 -Schematic diagram of the structure of the semantic integration module of the present invention.

[0043] Figure 6 -Visual comparison chart of the present invention and other COD detection models.

[0044] Figure 7 - Visual comparison of F-measure of the present invention and other COD detection models.

[0045] Figure 8 -Visual comparison chart of the present invention and other COD detection models PR. DETAILED DESCRIPTION

[0046] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0047] The present invention is based on the camouflage target detection method of edge information and coarse segmentation, and the overall architecture of the model is as follows: Figure 1 As shown, the invention designs an innovative multi-module system to address the key challenges in camouflaged target detection. The detection model of the present invention includes four core components: edge perception module (EPM), feature enhancement module (FEM), coarse segmentation module (CSM) and semantic integration module (SIM). First, the design purpose of FEM is to capture and process features from different scales and receptive fields. Through this process, FEM can significantly enhance the effective feature information that is crucial to the recognition of camouflaged targets, providing a richer and more accurate data basis for subsequent modules. EPM focuses on combining features from high and low layers to generate edge auxiliary information that helps to identify camouflaged targets. This information is crucial to improving the accuracy of target detection, especially in target recognition tasks under complex backgrounds, where the role of EPM is particularly prominent. The task of CSM is to integrate features from multiple levels to generate a preliminary segmented image of the camouflaged target. This step is of great significance for quickly determining the approximate position and shape of the target, and provides preliminary guidance for the final precise segmentation. SIM combines the edge information generated by the previous modules with the coarse segmentation map, and further optimizes the detection performance of the entire system through deep information fusion technology. SIM ensures that the final output not only accurately locates and identifies the camouflaged target, but also provides high-quality segmentation results to meet the needs of practical applications. In order to verify the effectiveness and practicality of the present invention, the present invention conducts extensive experiments on six different datasets and conducts detailed visualization analysis of the experimental results.

[0048] 1. Feature Enhancement Module (FEM)

[0049] When the human visual system observes a camouflaged image with a highly similar foreground and background, it will gradually focus and gradually narrow the observation area according to the complexity of the image until the foreground position is accurately determined. Since the input camouflaged image contains a large number of effective target features and interfering background noise, relying solely on the basic feature extraction of the backbone network is not enough to effectively filter out these interferences. Therefore, this method considers introducing a feature enhancement module (FEM) in the network. FEM consists of four sub-branches B k (k=1, 2, 3, 4) and a main branch B O Each sub-branch consists of a depth-wise separable convolution, such as Figure 2As shown in FIG. 1 , the FEM architecture design proposed in the present invention introduces a large-kernel depthwise separable convolution in the feature enhancement module to expand the receptive field and achieve more efficient feature capture. Assume that the input feature of the feature enhancement module is f i (i=1, 2, 3, 4, 5), the output of the feature enhancement module can be expressed as:

[0050]

[0051] 2. Edge Perception Module (EPM)

[0052] Effective edge priors not only play a key role in target localization tasks, but also provide important support in image segmentation, which helps to accurately divide the boundaries of foreground and background, thereby significantly improving the accuracy and reliability of detection. This has been confirmed by numerous studies. Although low-level features contain rich edge details, they also introduce a large amount of non-target edge information, which may lead to serious errors in subsequent detection tasks. Therefore, it is necessary to use high-level and low-level semantic or position information to assist in the extraction of edge features related to camouflage. Inspired by these studies, the present invention proposes an edge-aware module (EPM) to extract edge information more effectively. The EPM architecture proposed in the present invention is used to fuse high-level and low-level features to generate auxiliary edge information. In EPM, we combine low-level and high-level features (f2~f5) to model target-related edge information, such as Figure 3 Specifically, a 1×1 convolutional layer is used to adjust the number of channels to 64, and the features (f3 to f5) are upsampled to match the dimension of f2, and channel concatenation is performed. Then, feature maps of adjacent scales are concatenated to preserve multi-scale information. Where i, j represent feature maps of adjacent scales, and ⊕ represents a concatenation operation. In addition, the module structure contains multiple feature filter blocks (FFBs) to reduce noise, occlusion, and other interference in the image, thereby maximizing the effectiveness of edge information extraction.

[0053] FFB is the core component of EPM, which is specially used to process boundary-related information. Figure 3 As shown in b, FFB can flexibly adjust the receptive field of the feature map. Specifically, the present invention concatenates the feature map with the features of its adjacent scales to form a 23 For example, the concatenated feature f 23 The processing is done through two independent branches, one of which consists of a 1×1 convolutional layer and a sigmoid activation function. Consider the input tensor f 23 =[f 1，1 , f 2，2 , ...f i，j , ...f H，W ] is a slice representation, where Represents the local features at the spatial position (i, j), i∈{1, 2, ..., H}, j∈{1, 2, ..., W}. The convolution operation is used to perform spatial compression and generate a projection tensor q∈R H×W This projection tensor is adjusted to the range [0,1] by the Sigmoid layer σ(·), generating the activation of the spatial region The other branch consists of a 1×1 convolutional layer followed by two 3×3 convolutional layers, each followed by batch normalization and ReLU activation layers to generate the output feature map f′ 23 ∈R H ×W×2 Then the Softmax function is applied to the feature map to obtain two weight vectors W2 and W3, which are used to fuse the feature map. The expression is Therefore, the final output f″ 23 It can be expressed as:

[0054] f″ 23 =U+V#(2)

[0055] in, represents element-wise multiplication. The other branches follow a similar process (e.g., f 34 ). Finally, the multi-scale feature maps are integrated through residual connections, and the expression is as follows:

[0056]

[0057] 3. Coarse Segmentation Module (CSM)

[0058] like Figure 4 As shown in the figure, CSM is used to generate a rough segmentation map of the disguised target. This module takes multi-level feature maps f1, f2, f3, f4 and f5 as input, and gradually fuses shallow features into deep features through FEM. This fusion process not only retains the rich semantic information in the deep features, but also enhances the detail representation ability of the shallow features. Finally, the fused features are processed by a 3×3 convolutional layer to generate a rough segmentation output. This rough segmentation map provides preliminary information about the location and shape of the disguised target, guiding the subsequent fine segmentation stage. The detailed processing process of CSM is described as follows:

[0059]

[0060] 4. Semantic Integration Module (SIM)

[0061] In order to achieve more accurate prediction, the present invention integrates multi-level fusion features and designs SIM, which aims to combine edge information with coarse segmentation maps to enhance contextual semantics and effectively detect disguised targets. Different from the global context module in BBSNet, which only uses several independent parallel branches to extract features of different scales without considering the semantic connection between the branches, the present invention's SIM improves contextual semantic information through cross-scale fusion. Figure 5 As shown, the present invention extracts edge information And the generated rough segmentation map Combine and incorporate useful information into feature f′ i To obtain the final segmentation map. f′ i The processing of the three pieces of information is described as follows:

[0062]

[0063] When the input of the high-level semantic integration module needs to be combined, the output of the low-level semantic integration module is expressed as follows:

[0064]

[0065] 5. Loss Function

[0066] In order to maximize the distinction between foreground and background, the present invention designs a loss function consisting of two parts: prediction loss L p and marginal loss L e For the prediction loss L p , this paper introduces weighted binary cross entropy loss and weighted intersection-over-union loss These loss functions focus on the representation of key pixels by calculating and considering the differences in the center pixels, rather than giving equal weight to all pixels. The specific description is as follows:

[0067]

[0068] Among them, G t The mask of the camouflaged target, P i (i=2, 3, 4, 5) represents four camouflaged target masks generated by EPM.

[0069] To handle the edge loss L e , the present invention uses Dice loss L dice to automatically handle the significant imbalance between positive and negative samples. Therefore, the total loss function can be expressed as follows:

[0070] L total =L p +λLdice (P e , G e )#(7)

[0071] Among them, λ is a hyperparameter used to balance the relationship between object edge loss and overall loss. In the embodiment of the present invention, λ is set to 3. e represents the edge of the camouflaged target, and P e Represents the prediction of the edge of the camouflaged target.

[0072] The present invention is mainly achieved through the following specific technical solutions:

[0073] S1: Prepare and preprocess training data sets; the present invention collects rich data sets, including CHAMELEON, CAMO, COD10K, NC4K, ACOD2K and CAMOUFLAGE.

[0074] S2: A camouflaged target detection model is designed according to the model architecture provided by the present invention, and the four main modules are EPM, FEM, CSM and SIM.

[0075] S3: Design a loss function for learning according to the loss function designed by the present invention.

[0076] S4: Send the preprocessed data set to the disguised target detection model for training, and save the model with the best training effect as the final detection model.

[0077] S5: After the model training is completed, the model is tested and the effectiveness of the model is verified using public evaluation indicators.

[0078] S6: Input the RGB image to be detected into the disguised target detection model, and the output of the disguised target detection model is the disguised target image obtained by detection.

[0079] Among them, the specific method of preprocessing in S1 is: use the Adam optimizer to pre-train the diagnostic model, set the learning rate to 1e-4, and the batch size to 16. In order to reduce the randomness in the detection results, the present invention conducted 10 experiments under different random initialization conditions, and averaged the results of these experiments to obtain the final diagnostic results. In addition, in order to ensure the consistency of the experiments, all calculations were performed in the same environment, which included Intel Xeon (R) Gold 6226R CPU, NVIDIA Quadro RTX 5000 GPU, Windows 10 operating system, CUDA 11.2 and PyTorch 2.3.0. During training and testing, all input images were adjusted to a size of 416×416. In order to prevent the model from overfitting, data enhancement techniques such as rotation, random horizontal flipping, random cropping and edge cropping were used.

[0080] The specific approach to designing the model architecture and four modules in S2 is as follows: Taking RGB images as input, Res2Net-50 is used as the backbone network. Initially, the backbone network extracts preliminary features from each layer, denoted as f i (i=1, 2, 3, 4, 5). Input features f2~f5 into EPM, and the edge labels guide the effective extraction of edge information of the camouflaged target from each layer. The generated edge detection feature map is expressed as At the same time, features f1~f5 are input into FEM, which removes redundant noise, expands the receptive field, and reduces the number of channels to 64, further refining the initial features extracted by the backbone network. Next, we designed CSM to generate a rough map using local feature information. Finally, SIM combines the high-level and low-level features of each layer, and sequentially uses the output of high-level SIM as the input of low-level SIM to finally generate a prediction map.

[0081] The specific method of S3 loss function is as follows: In order to maximize the distinction between foreground and background, a loss function consisting of two parts is designed: the prediction loss L p and marginal loss L e In order to deal with the edge loss L e , using Dice loss L dice to automatically handle the significant imbalance between positive and negative samples.

[0082] The specific method of training the preprocessed data in S4 is as follows: During the training and testing process, all input images are adjusted to a size of 416 × 416. In order to prevent the model from overfitting, data enhancement techniques such as rotation, random horizontal flipping, random cropping and edge cropping are used.

[0083] The specific method of verifying the effectiveness of the model in S5 is: using the round parameters with the best training effect, sending the test set in the data set to the model to calculate the best index. The effectiveness of the model is verified using the following four evaluation indicators.

[0084] (1) Mean absolute error (MAE):

[0085]

[0086] (2) Evaluating regional perception S r and object perception S o The structural similarity between m ):

[0087]

[0088] (3) Combining the effects of precision and recall through a weighted sum (F β ):

[0089]

[0090] (4) Combining the local pixel value with the image-level average value (E m ):

[0091]

[0092] Compared with other methods, the proposed method shows strong competitiveness in six data sets. Specifically, in the Camouflage data set, the proposed method has m 、F β and E m The results in MAE, S m 、F β and E m The performance of the proposed method is 0.073, 0.791, 0.687 and 0.843 respectively, second only to TJNet. On the NC4K dataset, due to the complex and diverse shapes of the camouflaged objects and the influence of edge uncertainty, detecting camouflaged objects becomes more challenging. Most methods have large errors on this dataset. m 、F β and E m The results are 0.044, 0.861, 0.794 and 0.908 respectively. Compared with the previous best method, the proposed method improves MAE by 23.6%, F β The improvement is 10.1%. This shows that the proposed method achieves significant performance improvement in dealing with severe occlusion and multiple targets.

[0093] Table 1 is a comparison of the test results of the proposed method and other methods on 6 public data sets (bold indicates the best, underline indicates the second best, and dotted line indicates the third best);

[0094]

[0095] Figure 6 This is a visual comparison diagram of the present invention and other COD detection models. Figure 7 This is a visual comparison diagram of F-measure of the present invention and other COD detection models. Figure 8 This is a visual comparison chart of the present invention and other COD detection models PR.

[0096] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the applicant has described the present invention in detail with reference to the preferred embodiments, ordinary technicians in the field should understand that modifications or equivalent substitutions of the technical solution of the present invention without departing from the purpose and scope of the technical solution should be included in the scope of the claims of the present invention.

Claims

1. A camouflaged target detection method based on edge information and coarse segmentation, characterized in that the steps are as follows: 1) Training a disguised target detection model; the disguised target detection model includes a backbone network, an edge perception module, a feature enhancement module, a rough segmentation module and a semantic integration module; 1.1) The original RGB image dataset is fed into a predefined backbone network; the backbone network consists of a series of convolutional layers that decompose the input image into five different resolution levels by gradually downsampling it; at each level, preliminary features of different levels are extracted from the image through the application of convolution operations and activation functions, denoted as f i (i=1, 2, 3, 4, 5); 1.2) Input features f2~f5 into the edge perception module. The edge labels guide the effective extraction of edge information of the camouflaged target from each layer. The generated edge detection feature map is expressed as At the same time, features f1~f5 are respectively input into the feature enhancement module, through which redundant noise is removed, the receptive field is expanded, and the number of channels is reduced to 64, and the initial features extracted by the backbone network are further refined. The output of the feature enhancement module is expressed as f′ i (i=1,2,3,4,5); 1.3) The coarse segmentation module is used to generate a rough segmentation map of the camouflaged target; the coarse segmentation module takes the multi-level feature maps f1, f2, f3, f4 and f5 as input, and gradually fuses the deep features into the shallow features through the feature enhancement module, and processes the fused features through a 3×3 convolution layer to generate a coarse segmentation map, which is expressed as 1.4) The edge detection feature map generated by the edge perception module The output f′ of the feature enhancement module i (i=2,3,4,5) and the coarse segmentation map generated by the coarse segmentation module The corresponding input is sent to the semantic integration module; the semantic integration module combines the high-level and low-level features of each layer to output the corresponding prediction graphs respectively, and then the prediction graphs output by the high-level semantic integration module are used as the inputs of the low-level semantic integration modules in turn. The output of the lowest-level semantic integration module is the final prediction graph, which is the output of the disguised target detection model; 2) The RGB image to be detected is input into the disguised target detection model, and the output of the disguised target detection model is the disguised target image obtained by detection.

2. The method for detecting camouflaged targets based on edge information and coarse segmentation according to claim 1, characterized in that: In step 1.2), the feature enhancement module consists of four sub-branches B k (k=1, 2, 3, 4) and a main branch B O Each sub-branch contains a depth-wise separable convolution, and the input of the feature enhancement module is f i (i=1, 2, 3, 4, 5), the output is expressed as:

3. The method for detecting camouflaged targets based on edge information and rough segmentation according to claim 1, characterized in that: In step 1.2), the edge perception module is used to fuse high-level and low-level features to generate auxiliary edge information; in the edge perception module, a 1×1 convolution is used to adjust the number of channels to 64, and the features (f3 to f5) are upsampled to match the dimension of f2, and channel splicing is performed; then, features of adjacent scales are spliced ​​to retain multi-scale information.

4. The method for detecting camouflaged targets based on edge information and rough segmentation according to claim 3, characterized in that: The edge perception module contains multiple feature filtering blocks, which are used to transform the feature f m and the feature f of its adjacent scale n Splice and get f mn ; The concatenated feature f mn It is processed by two independent branches, one of which includes a 1×1 convolutional layer and a sigmoid activation function; considering the input tensor f mn =[f 1，1 , f 2，2 , ..., f i，j , ..., f H，W ] is a slice representation, where Represents the local features at the spatial position (i, j), i∈{1, 2, ..., H}, j∈{1, 2, ..., W}; performs spatial compression through convolution operations to generate a projection tensor q∈R H×W ; The projection tensor is adjusted to the range [0,1] through the Sigmoid layer σ(·) to generate the activation of the spatial region The other branch consists of a 1×1 convolutional layer followed by two 3×3 convolutional layers; each convolutional layer is followed by batch normalization and ReLU activation layers to generate the output feature map f′ mn ∈R H×W×2 ; Then apply the Softmax function to the feature map to obtain two weight vectors W m and W n , used to fuse feature maps, the expression is The final output f″ mn It is expressed as: f″ mn =U+V#(2) in, Represents element multiplication; Finally, the multi-scale feature maps are integrated through residual connections, and the expression is as follows:

5. The method for detecting camouflaged targets based on edge information and rough segmentation according to claim 1, characterized in that: In step 1.3), the specific processing process of the rough segmentation module is described as follows:

6. The method for detecting camouflaged targets based on edge information and rough segmentation according to claim 1, characterized in that: In step 1.4), the semantic integration module will detect the feature map from the edge and the rough segmentation map The useful information extracted from the feature enhancement module is integrated into the output feature f′ i (i=2,3,4,5), the processing process is described as follows: Combined with the input of the high-level semantic integration module, the output of the low-level semantic integration module is expressed as follows:

7. The method for detecting camouflaged targets based on edge information and rough segmentation according to claim 1, characterized in that: Step 1) When training the camouflaged target detection model, the distinction between foreground and background is improved through the loss function, which includes the prediction loss L p and marginal loss L e ; Prediction loss L p It is expressed as follows: in, is the weighted binary cross entropy loss; is the weighted intersection-over-union loss; G t The mask of the camouflaged target, P i (i=2, 3.4, 5) represents four camouflaged target masks generated by the edge-aware module; Margin loss L uses Dice loss L dice To automatically handle the significant imbalance between positive and negative samples; The total loss function is expressed as follows: THE total =L p +λL dice (P e ,G e )#(7) Among them, λ is a hyperparameter used to balance the relationship between object edge loss and overall loss; L dice is the Dice loss; G e represents the edge of the camouflaged target, P e Represents the prediction of the edge of the camouflaged target.

Images