A camouflage target detection method based on edge information and coarse segmentation

Abstract

The application discloses a camouflage target detection method based on edge information and coarse segmentation, extracts preliminary features of an image from each layer of a backbone network and records the features as f 1 ~f 5 ; features f 2 ~f 5 are input into an edge perception module, edge labels guide effective extraction of edge information of the camouflage target from each layer to generate an edge feature map. f 1 ~f 5 are input into a feature enhancement module to further refine the preliminary features extracted by the backbone network. f 1 ~ f 5 as input, gradually fuse deep features into shallow features through the feature enhancement module, process the fused features through convolution to generate a coarse segmentation map. A semantic integration module integrates useful information extracted from the edge feature map and the coarse segmentation map into the output features of the feature enhancement module to finally obtain a prediction map. The application can significantly improve the segmentation accuracy and detail performance, and can more accurately recognize and segment the camouflage target in a complex background.

Description

Technical Field

[0001] This invention relates to target detection, specifically to a method for detecting camouflaged targets based on edge information and coarse segmentation. This invention can be applied to the detection of camouflaged targets in autonomous driving, intelligent security, and military reconnaissance, so as to improve the accuracy and response efficiency of identifying concealed or camouflaged objects, and belongs to the field of intelligent recognition technology. Background Technology

[0002] Camouflage is a strategy used by organisms in nature to conceal themselves by blending closely with their surroundings. Due to their high similarity to the background, they are often difficult for predators and prey to spot. The task of camouflage target detection (COD) is to segment targets that are difficult for the human eye to identify, especially those that are closely integrated with complex backgrounds. COD technology has been widely applied 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), thus attracting increasing attention.

[0003] Early COD methods relied heavily on handcrafted low-level features, which were easily affected by complex backgrounds, resulting in low detection accuracy. For example, traditional image processing techniques might fail to effectively separate the target from the background if the texture, color, and shape were too similar to the target. To address this issue, many deep learning-based COD methods have emerged. These methods leverage the powerful representational capabilities of neural networks, automatically learning more complex feature patterns from large amounts of data, demonstrating significant potential. Current COD methods perform well in detecting single camouflaged targets, accurately identifying them in relatively simple scenes. However, their performance degrades significantly when faced with complex scenes involving multiple targets, occlusion, and overlapping objects. Especially when the foreground and background are highly similar, even advanced algorithms like TPRNet and C2FNet fail to accurately reproduce the complete structure of camouflaged targets when attempting to detect them within camouflaged regions. This is primarily because these advanced methods fail to fully recognize the importance of edge information in camouflaged target detection. Edge information provides crucial clues about the target's contour, helping the algorithm more accurately define the target's boundaries. Furthermore, advanced methods like EAMNet have limitations in handling camouflaged targets with slight occlusion, leading to both false positives and false negatives, ultimately resulting in inaccurate prediction maps. This is primarily due to inaccurate edge prior guidance and insufficient position prediction; the algorithm fails to effectively utilize existing edge information to guide target localization and segmentation. Therefore, an ideal camouflaged target detection method should utilize more representative features from multiple perspectives to compensate for the limitations of human vision.

[0004] Despite the achievements of existing camouflage target detection technologies, significant challenges remain in several key areas. First, the high similarity between camouflage targets and the background leads to insufficient detection capabilities against complex backgrounds, especially in natural environments where the diversity and complexity of backgrounds further exacerbate this problem. Second, existing technologies have limited processing capabilities for partially occluded or overlapping camouflage targets, frequently resulting in detection failures or incomplete segmentation. Furthermore, edge information plays a crucial role in camouflage target detection, but many existing methods fail to fully exploit or effectively utilize this information, leading to less than ideal detection results. To address these issues, future research needs to focus on developing more efficient feature extraction and representation methods to improve the robustness of models to complex backgrounds and the accuracy of multi-target detection; simultaneously, 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 to reduce training time will help promote the development and application of camouflage target detection technologies. Summary of the Invention

[0005] To address the problems of high training costs and narrow applicability in most existing camouflage target detection methods, especially the segmentation difficulties caused by large shape differences, blurred boundaries, and small target segmentation objects among different camouflage targets, the purpose of this invention is to provide a camouflage target detection method based on edge information and coarse segmentation. This invention can significantly improve the segmentation accuracy and detail representation, and can more accurately identify and segment camouflage targets in complex backgrounds and multi-target scenes, while also providing fast and efficient target recognition.

[0006] The technical solution of this invention is implemented as follows:

[0007] A method for detecting camouflaged targets based on edge information and coarse segmentation, comprising the following steps:

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

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

[0010] 1.2) Input features f2 to f5 into the edge perception module. The edge labels guide the extraction of edge information of the camouflaged target from each layer, generating an edge detection feature map, denoted as f. i e , where i = 2, 3, 4, 5;

[0011] Simultaneously, features f1 to f5 are input into the feature enhancement module. The feature enhancement module 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. The output of the feature enhancement module is represented as f. i ′, where i = 1, 2, 3, 4, 5;

[0012] 1.3) The coarse segmentation module is used to generate a preliminary segmentation map of the camouflaged target. The coarse segmentation module takes 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. The fused features are then processed by a 3×3 convolutional layer to generate the coarse segmentation map, denoted as f. i m , where i = 2, 3, 4, 5;

[0013] 1.4) The edge detection feature map f generated by the edge perception module i e The output f of the feature enhancement module i The coarse segmentation map f generated by the coarse segmentation module and the coarse segmentation module i m The corresponding input is given 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 map, and then uses the prediction map output by the high-level semantic integration module as the input of the low-level semantic integration module. The output of the lowest-level semantic integration module is the final prediction map, which is the output of the camouflage target detection model.

[0014] 2) Input the RGB image to be detected into the camouflage target detection model, and the output of the camouflage target detection model is the detected camouflage target image.

[0015] Furthermore, in step 1.2), the feature enhancement module consists of four sub-branches B k Where k = 1, 2, 3, 4; and a main branch B. O The system is composed of sub-branches, each containing depthwise separable convolutions, with the input to the feature enhancement module being f. i Where i = 1, 2, 3, 4, 5; the output is represented as:

[0016]

[0017] Further, 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 features (f3~f5) are upsampled to match the dimension of f2, and channel concatenation is performed; then, features of adjacent scales are concatenated to preserve multi-scale information.

[0018] Furthermore, the edge-aware module contains multiple feature filtering blocks, which are used to filter features f m Features f of adjacent scales n By splicing, we get f mn The spliced ​​features f mn The process is performed through two independent branches, one of which includes a 1×1 convolutional layer and a sigmoid activation function; consider the input tensor f. mn =[f 1,1 ,f 2 ,2 ,...,f i,j ,...,f H,W The slice representation of ], where Let represent the local features at spatial location (i, j), where i ∈ {1, 2, ..., H} and j ∈ {1, 2, ..., W}; spatial compression is performed through convolution to generate the 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 a batch normalization and ReLU activation layer to generate an output feature map f′. mn ∈R H×W×2 Then, the Softmax function is applied to the feature map to obtain two weight vectors W. m and W n , used for fusing feature maps, the expression is The final output is f″ mn Represented as:

[0019] f″ mn =U+V (2)

[0020] in, This represents element-wise multiplication; finally, the multi-scale feature maps are integrated through residual connections, as shown in the following expression:

[0021]

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

[0023]

[0024] Further, in step 1.4), the semantic integration module will extract the edge detection feature map f i e and coarse segmentation image f i m The useful information extracted is incorporated into the output feature f of the feature enhancement module. i The processing procedure is described as follows:

[0025]

[0026] Combining the input from 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 camouflage target detection model, the discriminative power between the foreground and background is improved through a loss function, which includes a prediction loss L. p and edge loss L e Predicted loss L p It is expressed as follows:

[0029]

[0030] in, The weighted binary cross-entropy loss; For weighted average loss; G t P represents the mask used to spoof the target. i , where i = 2, 3, 4, 5 represent four disguised target masks generated by the edge perception module;

[0031] Edge loss L e Dice loss L dice To automatically handle significant imbalances 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 (8)

[0034] Where λ is a hyperparameter used to balance the relationship between the edge loss and the overall loss of the object; L dice For Dice loss; G e P represents the edge of the camouflaged target. e This indicates a prediction of the edge of a camouflaged target.

[0035] This invention proposes a novel, fast, and accurate camouflage segmentation network model. This network effectively guides the network to explore and utilize representative edge features, avoiding the loss of edge details. Simultaneously, coarse segmentation enables rapid searching within the camouflage target area, determining the approximate location of the camouflage target and allowing subsequent fine processing to focus on these areas, avoiding blind searching across the entire image. Finally, effectively combining edge information with the coarse segmentation map significantly improves segmentation accuracy and detail representation. By utilizing key contour features provided by edge information to supplement missing details in the coarse segmentation map, this network design alleviates the challenges associated with edge detail loss and incomplete segmentation, thereby improving detection accuracy.

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

[0037] Compared to existing COD (Covered Target Detection) technologies, this invention significantly improves segmentation accuracy and detail by effectively combining edge information with a coarse segmentation map, enabling more accurate identification and segmentation of camouflaged targets in complex backgrounds and multi-target scenes. Secondly, this combination method accelerates processing speed and reduces computational resource consumption, making the algorithm more efficient in real-time applications. Furthermore, this invention performs exceptionally well in handling segmented objects with blurred boundaries, large shape differences, and small targets, effectively solving common problems of incomplete segmentation and false detections in existing technologies, thus enhancing the algorithm's robustness and adaptability. In summary, this invention not only achieves a technological breakthrough but also demonstrates significant advantages in practical applications. Attached Figure Description

[0038] Figure 1 - A schematic diagram of the model architecture of the camouflage target detection method of the present invention.

[0039] Figure 2 - A schematic diagram of the feature enhancement module structure of this invention.

[0040] Figure 3 - A schematic diagram of the edge sensing module structure of this invention.

[0041] Figure 4 - A schematic diagram of the rough segmentation module structure of this invention.

[0042] Figure 5 - A schematic diagram of the semantic integration module structure of this invention.

[0043] Figure 6 - Visual comparison chart of this invention with other COD detection models.

[0044] Figure 7 - Visual comparison chart of this invention with other COD detection models, F-measure.

[0045] Figure 8 - A visualization comparison of this invention with other COD detection models. Detailed Implementation

[0046] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0047] This invention presents a camouflaged target detection method based on edge information and coarse segmentation. The overall model architecture is as follows: Figure 1 As shown, this invention designs an innovative multi-module system designed to address key challenges in camouflaged target detection. The detection model comprises four core components: an Edge Awareness Module (EPM), a Feature Enhancement Module (FEM), a Coarse Segmentation Module (CSM), and a Semantic Integration Module (SIM). First, the FEM is designed to capture and process features from different scales and receptive fields. Through this process, the FEM significantly enhances the effective feature information crucial for camouflaged target identification, providing a richer and more accurate data foundation for subsequent modules. The EPM focuses on combining features from high and low levels to generate edge-aiding information that aids in identifying camouflaged targets. This information is crucial for improving the accuracy of target detection, especially in target recognition tasks with complex backgrounds, where the EPM's role is particularly prominent. The CSM integrates features from multiple levels to generate a preliminary segmentation image of the camouflaged target. This step is significant for quickly determining the approximate location and shape of the target, providing initial guidance for the final accurate segmentation. The SIM combines the edge information generated by the previous modules with the coarse segmentation image, further optimizing the overall system's detection performance through deep information fusion technology. SIM ensures that the final output not only accurately locates and identifies camouflaged targets, but also provides high-quality segmentation results, thus meeting the needs of practical applications. To verify the effectiveness and practicality of this invention, extensive experiments were conducted on six different datasets, and the experimental results were visualized and analyzed in detail.

[0048] 1. Feature Enhancement Module (FEM)

[0049] When the human visual system observes a camouflaged image where the foreground and background are highly similar, it gradually focuses and narrows the observation area according to the complexity of the image until the foreground location 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 insufficient to effectively filter out these interferences. Therefore, this method considers introducing a Feature Enhancement Module (FEM) into the network. The FEM consists of four sub-branches B... k Where k = 1, 2, 3, 4 and a main branch B O The composition consists of each sub-branch containing depthwise separable convolutions, such as... Figure 2 As shown, the FEM architecture proposed in this invention introduces a large-kernel depthwise separable convolution in the feature enhancement module to expand the receptive field and achieve more efficient feature capture. Let f be the input feature of the feature enhancement module. i Where i = 1, 2, 3, 4, 5, the output of the feature enhancement module can be expressed as:

[0050]

[0051] 2. Edge Awareness Module (EPM)

[0052] Effective edge priors not only play a crucial role in target localization tasks but also provide important support in image segmentation, helping to accurately delineate foreground and background boundaries, thereby significantly improving detection accuracy and reliability. 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 utilize high-level and low-level semantic or positional information to assist in the extraction of camouflage-related edge features. Inspired by these studies, this invention proposes an Edge Aware Module (EPM) to extract edge information more effectively. The proposed EPM architecture is used to fuse high-level and low-level features to generate auxiliary edge information. In the EPM, we combine low-level and high-level features (f2–f5) to model target-related edge information, such as… Figure 3 As shown in diagram a. Specifically, a 1×1 convolutional layer is used to adjust the number of channels to 64, and features (f3–f5) are upsampled to match the dimension of f2, followed by channel concatenation. Then, feature maps from adjacent scales are concatenated to preserve multi-scale information. Where i and j represent feature maps at adjacent scales, and ⊕ represents the stitching operation. Furthermore, the module structure includes 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 a core component of EPM, specifically designed to handle boundary-related information. For example... Figure 3 As shown in b, FFB can flexibly adjust the receptive field of the feature map. Specifically, this invention concatenates the feature map with features of its adjacent scales, using f... 23 For example, the concatenated feature f 23 The process is performed through two separate 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 The slice representation of ], where Let q represent the local features at spatial location (i, j), where i ∈ {1, 2, ..., H} and j ∈ {1, 2, ..., W}. Spatial compression is performed through convolution to generate the projected tensor q ∈ R. H×W This projection tensor is adjusted to the range [0,1] by a 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 with a batch normalization and ReLU activation layer to generate an 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 maps. The expression is as follows: Therefore, the final output is f″ 23 It can be represented as:

[0054] f″ 23 =U+V#(2)

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

[0056]

[0057] 3. Coarse Segmentation Module (CSM)

[0058] like Figure 4 As shown, CSM is used to generate a preliminary segmentation map of the camouflaged target. This module takes multi-level feature maps f1, f2, f3, f4, and f5 as input and progressively fuses shallow features into deep features using FEM. This fusion process not only preserves the rich semantic information in the deep features but also enhances the detail representation capability of the shallow features. Finally, a 3×3 convolutional layer processes the fused features to generate a coarse segmentation output. This coarse segmentation map provides preliminary information about the location and shape of the camouflaged target, guiding the subsequent fine segmentation stage. The detailed processing procedure of CSM is described below:

[0059]

[0060] 4. Semantic Integration Module (SIM)

[0061] To achieve more accurate predictions, this invention integrates multi-level fusion features and designs a Simulated Context (SIM). Its purpose is to combine edge information with a coarse segmentation map to enhance contextual semantics and effectively detect camouflaged targets. Unlike the Global Context Module in BBSNet, which uses only a few independent parallel branches to extract features at different scales without considering the semantic relationships between branches, this invention's SIM improves contextual semantic information through cross-scale fusion. Figure 5 As shown, the present invention extracts edge information f i e Compared with the generated coarse segmentation map Combine and incorporate useful information into feature f i In the process of obtaining the final segmentation map, only f is integrated. i e f i m f i The processing procedure for the three pieces of information is described as follows:

[0062]

[0063] When it is necessary to combine the input of the higher-level semantic integration module, the output of the lower-level semantic integration module is expressed as follows:

[0064]

[0065] 5. Loss Function

[0066] To maximize the distinction between foreground and background, this invention designs a two-part loss function: prediction loss L. p and edge loss L e For the predicted loss L p This invention introduces a weighted binary cross-entropy loss. Weighted intersection and comparison loss These loss functions focus on the representation of key pixels by calculating and taking into account the differences in the center pixels, rather than assigning the same weight to all pixels. A detailed description follows:

[0067]

[0068] Among them, G t P represents the mask used to spoof the target. i , where i = 2, 3, 4, 5 represent the four camouflage target masks generated by EPM.

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

[0070] L total =L p +λL dice (P e G e (8) Wherein, λ is a hyperparameter used to balance the relationship between object edge loss and overall loss. In this embodiment of the invention, λ is set to 3. e This represents the edge of the camouflaged target, while P e This indicates a prediction of the edge of a camouflaged target.

[0071] This invention is mainly achieved through the following specific technical solutions:

[0072] S1: Prepare and preprocess the training dataset; This invention has collected a rich dataset, including CHAMELEON, CAMO, COD10K, NC4K, ACOD2K and CAMOUFLAGE.

[0073] S2: Design a camouflage target detection model according to the model architecture provided by this invention. The four main modules are EPM, FEM, CSM and SIM.

[0074] S3: The loss function designed according to the present invention is used for learning.

[0075] S4: Feed the preprocessed dataset into the camouflage target detection model for training, and save the model with the best training effect as the final detection model.

[0076] S5: After the model is trained, test the model and use publicly available evaluation metrics to verify the effectiveness of the model.

[0077] S6: Input the RGB image to be detected into the camouflage target detection model, and the output of the camouflage target detection model is the detected camouflage target image.

[0078] The preprocessing in S1 involves pre-training the diagnostic model using the Adam optimizer with a learning rate of 1e-4 and a batch size of 16. To reduce randomness in the detection results, 10 experiments were conducted under different random initialization conditions, and the results of these experiments were averaged to obtain the final diagnostic result. Furthermore, to ensure consistency, all computations were performed in the same environment, including an Intel Xeon(R) Gold6226R CPU, an NVIDIA Quadro RTX 5000 GPU, a Windows 10 operating system, CUDA 11.2, and PyTorch 2.3.0. During training and testing, all input images were resized to 416×416. To prevent overfitting, data augmentation techniques such as rotation, random horizontal flipping, random cropping, and edge cropping were employed.

[0079] The specific implementation of the model architecture and four modules in S2 is as follows: For the process of taking RGB images as input, Res2Net-50 is used as the backbone network. Initially, preliminary features are extracted from each layer through the backbone network, denoted as f. i Where i = 1, 2, 3, 4, 5. Features f2 to f5 are input into the EPM, and edge labels guide the effective extraction of edge information of the camouflaged target from each layer. The generated edge detection feature map is represented as f. i e Where i = 2, 3, 4, 5. Simultaneously, features f1 to f5 are input into the 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 a CSM to generate a coarse map f using local feature information. i m Where i = 2, 3, 4, 5. Finally, SIM combines the high-level and low-level features of each layer, and sequentially uses the output of the high-level SIM as the input of the low-level SIM to finally generate the prediction map.

[0080] The S3 loss function is implemented as follows: To maximize the distinction between the foreground and background, a two-part loss function is designed, with a prediction loss L. p and edge loss L e To handle edge loss L e Using Dice loss L dice To automatically handle significant imbalances between positive and negative samples.

[0081] The specific steps for training the preprocessed data in S4 are as follows: During training and testing, all input images are resized to 416×416. To prevent overfitting, data augmentation techniques such as rotation, random horizontal flipping, random cropping, and edge cropping are employed.

[0082] The specific steps for verifying model effectiveness in S5 are as follows: using the parameters from the training epoch with the best performance, the test set from the dataset is fed into the model to calculate the best metric. The effectiveness of the model is then verified using the following four evaluation metrics.

[0083] (1) Mean Absolute Error (MAE):

[0084]

[0085] (2) Assessing regional perception S r and object perception S o Structural similarity between them (S m ):

[0086]

[0087] (3) Combining the effects of precision and recall through weighted summation (F β ):

[0088]

[0089] (4) Combine local pixel values ​​with image-level average values ​​(E m ):

[0090]

[0091] Compared to other methods, the method proposed in this invention demonstrates strong competitiveness across six datasets. Specifically, in the Camouflage dataset, the method proposed in this invention performs better in MAE, S... m F β and E m The proposed method achieved scores of 0.009, 0.815, 0.744, and 0.867 respectively, outperforming previous methods. In the ACOD2K dataset, the proposed method achieved high scores in MAE and S... m F β and E m The performance metrics achieved were 0.073, 0.791, 0.687, and 0.843, respectively, second only to TJNet. On the NC4K dataset, detecting camouflaged objects becomes more challenging due to the complex and varied shapes of the camouflaged objects and the uncertainty of their edges. Most methods exhibit significant errors on this dataset. In the method proposed in this invention, MAE and S... m Fβ and E m The values ​​reached 0.044, 0.861, 0.794, and 0.908, respectively. Compared with the previous best method, the method proposed in this invention improves the MAE by 23.6%, and F... β This represents a 10.1% improvement. This demonstrates that the proposed method achieves significant performance improvements when handling cases with severe occlusion and multiple targets.

[0092] Table 1 shows the test results of the method proposed in this invention and other methods on 6 public datasets (bold indicates the best, underline indicates the second best, and dashed line indicates the third best).

[0093]

[0094] Figure 6 This is a visualization comparison of the present invention with other COD detection models. Figure 7 This is a visualization comparison of the present invention with other COD detection models, such as F-measure. Figure 8 This is a visualization comparison chart of the present invention with other COD detection models.

[0095] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the applicant has described the present invention in detail with reference to preferred embodiments, those skilled in the art should understand that any modifications or equivalent substitutions to the technical solutions of the present invention without departing from the spirit and scope of the present invention should be covered within the scope of the claims of the present invention.

Claims

1. A method for detecting camouflaged targets based on edge information and coarse segmentation, characterized in that: The steps are as follows: 1) Train a camouflaged target detection model; the camouflaged target detection model includes a backbone network, an edge perception module, a feature enhancement module, a coarse segmentation module, and a semantic integration module; 1.1) The original RGB image dataset is input 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 performing progressive downsampling operations. At each level, preliminary features at different levels are extracted from the image through convolution operations and the application of activation functions, denoted as f. i , where i = 1, 2, 3, 4, 5; 1.2) Input features f2 to f5 into the edge perception module. Edge labels guide the extraction of edge information of the camouflaged target from each layer, generating an edge detection feature map, represented as follows: Where i = 2, 3, 4, 5; Simultaneously, features f1 to f5 are input into the feature enhancement module. The feature enhancement module 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. The output of the feature enhancement module is represented as f. i ′, where i = 1, 2, 3, 4, 5; 1.3) The coarse segmentation module generates a preliminary segmentation map of the camouflaged target. The coarse segmentation module takes 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. The fused features are then processed by a 3×3 convolutional layer to generate the coarse segmentation map, represented as follows: Where i = 2, 3, 4, 5; 1.4) The edge detection feature map generated by the edge perception module The output f of the feature enhancement module i The coarse segmentation map generated by the coarse segmentation module and the coarse segmentation module. The corresponding input is given 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 map, and then uses the prediction map output by the high-level semantic integration module as the input of the low-level semantic integration module. The output of the lowest-level semantic integration module is the final prediction map, which is the output of the camouflage target detection model. 2) Input the RGB image to be detected into the camouflage target detection model, and the output of the camouflage target detection model is the detected camouflage target image; In step 1.3), the specific processing procedure of the coarse segmentation module is described as follows: In step 1.4), the semantic integration module will extract the edge detection feature map and coarse segmentation map The useful information extracted is incorporated into the output feature f of the feature enhancement module. i The processing procedure is described as follows: Combining the input from the high-level semantic integration module, the output of the low-level semantic integration module is expressed as follows:

2. The camouflaged target detection method 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 Where k = 1, 2, 3, 4 and a main branch B O The system is composed of sub-branches, each containing depthwise separable convolutions, with the input to the feature enhancement module being f. i Where i = 1, 2, 3, 4, 5; the output is represented as:

3. The camouflaged target detection method based on edge information and coarse 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 features f3 to f5 are upsampled to match the dimension of f2 and channel concatenation is performed. Then, features of adjacent scales are concatenated to preserve multi-scale information.

4. The camouflaged target detection method based on edge information and coarse segmentation according to claim 3, characterized in that: The edge sensing module contains multiple feature filtering blocks, which are used to filter features f m Features f of adjacent scales n By splicing, we get f mn The spliced ​​features f mn The process is performed through two independent branches, one of which includes a 1×1 convolutional layer and a sigmoid activation function; consider the input tensor f. mn =[f 1，1 f 2，2 , ..., f i，j , ..., f H，W The slice representation of ], where Representing local features at spatial location (i, j), i∈{1,2,...,H}, j∈{1,2,...,W}; spatial compression is performed through convolution 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 a batch normalization and ReLU activation layer to generate an output feature map f′. mn ∈R H×W×2 Then, the Softmax function is applied to the feature map to obtain two weight vectors W. m and W n , used to fuse feature maps, the expression is Final output f″ mn Represented as: f″ mn =U+V (2) in, This represents element-wise multiplication; finally, the multi-scale feature maps are integrated through residual connections, as shown in the following expression:

5. The camouflaged target detection method based on edge information and coarse segmentation according to claim 1, characterized in that: Step 1) When training the camouflaged target detection model, the discrimination between the foreground and background is improved through a loss function, which includes the prediction loss L. p and edge loss L e Predicted loss L p It is expressed as follows: in, The weighted binary cross-entropy loss; For weighted average loss; G t P represents the mask used to spoof the target. i , where i = 2, 3, 4, 5 represent four disguised target masks generated by the edge perception module; Edge loss L e Dice loss L dice To automatically handle significant imbalances between positive and negative samples; The total loss function is expressed as follows: THE total =L p +λL dice (P e ,G e ) (8) Where λ is a hyperparameter used to balance the relationship between the edge loss and the overall loss of the object; L dice For Dice loss; G e P represents the edge of the camouflaged target. e This indicates a prediction of the edge of a camouflaged target.

Images