Camouflage target detection network fusing context awareness and background exploration

Through a camouflage object detection network that integrates context perception and background exploration, using Swin-Transformer and attention-linked context perception module AHCM and background exploration module BEM, the poor detection problem caused by camouflage objects being similar to the background is solved, and more accurate camouflage object and edge feature extraction.

CN119006844APending Publication Date: 2024-11-22CHINA THREE GORGES UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411153007.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-21
Publication Date
2024-11-22

AI Technical Summary

Technical Problem

Existing camouflage object detection technology cannot detect more complete camouflage objects and more accurate edge details when faced with camouflage objects highly similar to the background, especially when the edges are blurred.

Method used

Using a camouflage object detection network that integrates context perception and background exploration, the multi-scale feature extraction module Swin-Transformer, attention-concatenated context perception module AHCM and the background exploration module BEM of the fusion attention mechanism are used to hybrid convolution through multi-branches Patterns and dual attention mechanisms extract and integrate features to enhance the positioning of camouflage objects and the extraction of edge features.

Benefits of technology

It improves the detection accuracy and edge positioning accuracy of camouflage objects, can better distinguish camouflage objects and backgrounds, enhances the ability to extract edge features of camouflage objects, and significantly improves detection performance.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119006844A_ABST
    Figure CN119006844A_ABST
Patent Text Reader

Abstract

The invention discloses a camouflage target detection network fusing context sensing and background exploration. The network comprises a multi-scale feature extraction module Swinin-Transform, an attention cascade context sensing module AHCM, a full-connection decoder module FCD and a background exploration module BEM fusing an attention mechanism. The output of the multi-scale feature extraction module Swindow-Transform is connected with the input of the attention cascade context sensing module AHCM, and the output of the attention cascade context sensing module AHCM is connected with the input of the full-connection decoder module FCD and the input of the background exploration module BEM fused with the attention mechanism; the output of the full-connection decoder module FCD is connected with the input of the background exploration module BEM fused with the attention mechanism, and the output of the background exploration module BEM fused with the attention mechanism is used for outputting final prediction.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of computer technology, in particular to the field of target detection technology, and specifically to a technology for detecting camouflaged targets by fusing context perception and background exploration. Background Art

[0002] Camouflage is a common defense mechanism in nature. Some species use their appearance similar to the surrounding environment to protect themselves from predators. Camouflaged object detection (COD) aims to detect camouflaged objects hidden in the background. In recent years, COD technology has received more and more attention in many fields such as medicine (such as lung infection segmentation), industry (such as detection of surface defects of objects), agriculture (such as locust detection), and art (such as landscape image detection). It has strongly promoted the intelligent development of these fields and shown good development prospects.

[0003] Traditional COD methods mainly rely on low-level features based on manual design, such as direct visual features such as color, texture, and optical flow. When faced with complex scenes with extremely low contrast between foreground and background, there are usually problems such as time-consuming manual feature extraction, poor transferability, and low detection performance. In order to solve these problems, researchers have proposed a variety of camouflage target detection models based on deep learning in recent years. The target search and recognition network (SINet) proposed by Fan et al. expands the receptive field through the search module to obtain richer contextual information, and uses the recognition module to eliminate the interference of irrelevant features. Subsequently, the SINetV2 model they proposed mines the edge information of the camouflage target through the group-guided inversion attention module to generate fine predictions, but there may be uneven distribution of feature representations in the process of group guidance. Mei et al. constructed a COD network PFNet consisting of a positioning module and a focusing module. The network first locates potential targets from a global perspective, and then gradually predicts the camouflage target by focusing on the fuzzy area, but the direct multiplication of the current feature and the prediction prior during the focusing process may lead to feature confusion. Pang et al. designed a mixed-scale triple network ZoomNet, which uses a triple structure to extract features, then uses scale integration units to filter and aggregate features of a specific scale, and uses hierarchical mixed-scale units to further enhance the mixed-scale features, but this multivariate information strategy greatly increases the computational complexity of the network. Although the above methods have made some progress in the field of COD, due to the high similarity between the camouflaged object and the background and the blurring at the edge, it is impossible to detect more complete camouflaged objects and more accurate edge details. Summary of the invention

[0004] The purpose of the present invention is to solve the technical problem that the existing camouflaged target detection technology cannot detect a more complete camouflaged object and more accurate edge details when the camouflaged object is highly similar to the background, especially when the edges are blurred. A camouflaged target detection network that integrates context perception and background exploration is proposed.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0006] A camouflaged target detection network integrating context perception and background exploration. The camouflaged target detection network CABENet integrating context perception and background exploration includes a multi-scale feature extraction module Swin-Transformer, an attention cascade context perception module AHCM, a fully connected decoder module FCD, and a background exploration module BEM integrating an attention mechanism.

[0007] The output of the multi-scale feature extraction module Swin-Transformer is connected to the input of the attention cascade context perception module AHCM, the output of the attention cascade context perception module AHCM is connected to the fully connected decoder module FCD and the input of the background exploration module BEM fused with the attention mechanism, the output of the fully connected decoder module FCD is connected to the input of the background exploration module BEM fused with the attention mechanism, and the output of the background exploration module BEM fused with the attention mechanism is used to output the final prediction.

[0008] The multi-scale feature extraction module Swin-Transformer is used as the feature extraction layer to output multi-scale features O i , i = {1, 2, 3, 4}, features O2, O3, and O4 are fused through the feature pyramid to output features O′2, O′3, and O′4;

[0009] The feature O′2 is input to the first attention cascade context perception module AHCM, and the output of the first attention cascade context perception module AHCM is connected to the input of the fully connected decoder module FCD and the input of the background exploration module BEM of the first fusion attention mechanism;

[0010] The feature O′3 is input to the second attention cascade context perception module AHCM, and the output of the second attention cascade context perception module AHCM is connected to the input of the fully connected decoder module FCD and the input of the background exploration module BEM of the second fusion attention mechanism;

[0011] The feature O′4 is input to the third attention cascade context perception module AHCM, and the output of the third attention cascade context perception module AHCM is connected to the input of the fully connected decoder module FCD and the input of the background exploration module BEM of the third fusion attention mechanism;

[0012] The fully connected decoder module FCD outputs feature C5, which performs a reverse operation to output the edge prior value. Marginal Priors The input is given to the background exploration module BEM of the third fusion attention mechanism, and the output of the background exploration module BEM of the third fusion attention mechanism is pixel-added with the feature C5 to output the feature C4;

[0013] Feature C4 is input to the background exploration module BEM of the second fusion attention mechanism, and the output of the background exploration module BEM of the second fusion attention mechanism is pixel-added with feature C4 to output feature C3;

[0014] Feature C3 is input to the background exploration module BEM of the first fusion attention mechanism, and the output of the background exploration module BEM of the first fusion attention mechanism is pixel-added with feature C3 to output the final feature.

[0015] The attention cascade context perception module AHCM is specifically:

[0016] Any input of feature O'2, feature O'3, feature O'4 is given to the shortcut branch 4 parallel residual branches {b i ,i=1,2,3,4};

[0017] Quick branch It includes a first convolutional layer and a first spatial attention module PAM, the feature input is given to the first convolutional layer, and the output of the first convolutional layer is connected to the input of the first spatial attention module PAM;

[0018] The first residual branch b1 includes the first convolutional layer and the second spatial attention module PAM. The feature input is given to the first convolutional layer, and the output of the first convolutional layer is connected to the input of the second spatial attention module PAM.

[0019] The last three residual branches {b i , i=2,3,4} includes the first convolution layer, the second convolution layer, the third convolution layer with a specific expansion rate and the channel attention module CAM, the feature input is given to the first convolution layer, the output of the first convolution layer is connected to the input of the second convolution layer, the output of the second convolution layer is connected to the input of the third convolution layer with a specific expansion rate, and the output of the third convolution layer with a specific expansion rate is connected to the input of the channel attention module CAM;

[0020] The output of the channel attention module CAM is concatenated with the output of the second spatial attention module PAM, and the output after the concatenation is added to the output of the first spatial attention module PAM to output the enhanced features. feature And features

[0021] The background exploration module BEM that integrates the attention mechanism is as follows:

[0022] feature Input the background exploration module BEM of the first fusion attention mechanism, features Input the background exploration module BEM of the second fusion attention mechanism, features Input the background exploration module BEM of the third fusion attention mechanism;

[0023] In the third background exploration module BEM fused with attention mechanism, the feature and marginal priors Input to the first background exploration module BA, the first background exploration module BA outputs the updated feature and the updated marginal priors feature and marginal priors Input to the second background exploration module BA, the second background exploration module BA outputs the updated feature and the updated marginal priors feature and marginal priors The input is given to the third background exploration module BA, and the third background exploration module BA outputs the updated edge prior value

[0024] In the background exploration module BEM of the second fusion attention mechanism, the feature and marginal priors Input to the first background exploration module BA, the first background exploration module BA outputs the updated feature and the updated marginal priors feature and marginal priors Input to the second background exploration module BA, the second background exploration module BA outputs the updated feature and the updated marginal priors feature and marginal priors The input is given to the third background exploration module BA, and the third background exploration module BA outputs the updated edge prior value

[0025] In the background exploration module BEM of the first fusion attention mechanism, the feature and marginal priors Input to the first background exploration module BA, the first background exploration module BA outputs the updated feature and the updated marginal priors feature and marginal priors Input to the second background exploration module BA, the second background exploration module BA outputs the updated feature and the updated marginal priors feature and marginal priors The input is given to the third background exploration module BA, and the third background exploration module BA outputs the updated edge prior value

[0026] The background exploration module BA is specifically:

[0027] Marginal Priors After duplication according to the number of channels and features Add and input to the first convolutional layer, which outputs the updated features The output of the first convolutional layer is connected to the input of the second convolutional layer. The output of the second convolutional layer is multiplied by the edge prior value after sigmoid activation and finally multiplied by the edge prior value. Output updated edge priors

[0028] The camouflaged target detection network CABENet, which integrates context perception and background exploration, adopts the following steps when in use:

[0029] Step 1: The disguised target image is input to the Swin-Transformer multi-scale feature extraction module to extract global context information at multiple scales and output feature O i , i = {1, 2, 3, 4}, features O2, O3, and O4 are fused through the feature pyramid to output features O′2, O′3, and O′4;

[0030] Step 2: Features O′2, O′3, and O′4 are input to the attention cascade context perception module AHCM, which uses a parallel multi-branch hybrid convolution mode to expand the receptive field, enhance the global context information features, and use spatial attention and channel attention mechanisms to improve the perception of the target area, and output the enhanced features feature feature

[0031] Step 3: Features feature feature Input to the fully connected decoder module FCD to locate the disguised object, output the output feature C5, and perform a reverse operation on feature C5 to output the edge prior value

[0032] Step 4: Features feature feature The edge prior values ​​are input into the background exploration module BEM of the fusion attention mechanism. The extracted edge prior values ​​are fused with the target features layer by layer using the self-attention mechanism. The edge features of the camouflaged object are extracted by focusing on the background area in the image and the final features are output.

[0033] The above steps are used to detect camouflaged targets.

[0034] Compared with the prior art, the present invention has the following technical effects:

[0035] 1) In order to solve the problem of poor detection results caused by texture similarity and boundary fuzziness of camouflaged objects in COD, the present invention proposes a camouflaged target detection method that integrates context perception and background exploration, which uses the rich clues provided by context information and background information to distinguish camouflaged targets and backgrounds, and better locate camouflaged objects;

[0036] 2) The model of the present invention uses swin-transformer as the network skeleton to extract image features, obtains enhanced context feature representation through the AHCM module, uses FCD to effectively locate hidden objects, and finally uses the BEM module to make the model focus on the background area, better understand the boundary structure of the target, and improve the accuracy of boundary positioning;

[0037] 3) The present invention proposes an AHCM module, which expands the receptive field through different branches and obtains the multi-scale contextual information of the target. Then, through a dual attention mechanism, namely spatial attention and channel attention, the model's perception of details is improved, where the spatial attention mechanism is used to capture the importance of different areas in the image, and the channel attention mechanism is used to adjust the weights of feature channels to improve the importance of different channels; finally, the results of different branches are merged to enhance the model's perception of global information and strengthen the network's feature extraction capability.

[0038] 4) The present invention proposes a BEM module, which uses a self-attention mechanism to fuse the extracted edge prior with the target features layer by layer, and better understands the boundary structure of the target by focusing on the background area in the image, improves the accuracy of boundary positioning, and strengthens the extraction of edge features of camouflaged objects. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] The present invention will be further described below in conjunction with the accompanying drawings and embodiments:

[0040] Figure 1It is a framework diagram of the camouflage object detection network CABENet (Camouflage Object Detection method integrating context awareness and background exploration) in the present invention;

[0041] Figure 2 for Figure 1 Schematic diagram of the attention cascade context-aware module AHCM;

[0042] Figure 3 for Figure 1 Schematic diagram of the background exploration module (BEM) that integrates the attention mechanism;

[0043] Figure 4 is a comparison diagram of the model of the present invention and 6 other representative models;

[0044] Figure 5 This is a visualization result diagram of the ablation experiment in the present invention. DETAILED DESCRIPTION

[0045] like Figure 1 As shown, a camouflaged target detection network integrating context perception and background exploration, the network includes a multi-scale feature extraction module Swin-Transformer, an attention cascade context perception module AHCM, a fully connected decoder module FCD, and a background exploration module BEM integrating an attention mechanism;

[0046] The output of the multi-scale feature extraction module Swin-Transformer is connected to the input of the attention cascade context perception module AHCM, the output of the attention cascade context perception module AHCM is connected to the fully connected decoder module FCD and the input of the background exploration module BEM fused with the attention mechanism, the output of the fully connected decoder module FCD is connected to the input of the background exploration module BEM fused with the attention mechanism, and the output of the background exploration module BEM fused with the attention mechanism is used to output the final prediction.

[0047] The multi-scale feature extraction module Swin-Transformer is used as the feature extraction layer to output multi-scale features O i , i = {1, 2, 3, 4}, features O2, O3, and O4 are fused through the feature pyramid to output features O′2, O′3, and O′4;

[0048] The feature O′2 is input to the first attention cascade context perception module AHCM, and the output of the first attention cascade context perception module AHCM is connected to the input of the fully connected decoder module FCD and the input of the background exploration module BEM of the first fusion attention mechanism;

[0049] The feature O′3 is input to the second attention cascade context perception module AHCM, and the output of the second attention cascade context perception module AHCM is connected to the input of the fully connected decoder module FCD and the input of the background exploration module BEM of the second fusion attention mechanism;

[0050] The feature O′4 is input to the third attention cascade context perception module AHCM, and the output of the third attention cascade context perception module AHCM is connected to the input of the fully connected decoder module FCD and the input of the background exploration module BEM of the third fusion attention mechanism;

[0051] The fully connected decoder module FCD outputs feature C5, which performs a reverse operation to output the edge prior value. Marginal Priors The input is given to the background exploration module BEM of the third fusion attention mechanism, and the output of the background exploration module BEM of the third fusion attention mechanism is pixel-added with the feature C5 to output the feature C4;

[0052] Feature C4 is input to the background exploration module BEM of the second fusion attention mechanism, and the output of the background exploration module BEM of the second fusion attention mechanism is pixel-added with feature C4 to output feature C3;

[0053] Feature C3 is input to the background exploration module BEM of the first fusion attention mechanism, and the output of the background exploration module BEM of the first fusion attention mechanism is pixel-added with feature C3 to output the final feature.

[0054] like Figure 2 As shown in Figure 1, the attention cascade context perception module AHCM is specifically:

[0055] Any input of feature O'2, feature O'3, feature O'4 is given to the shortcut branch 4 parallel residual branches {b i ,i=1,2,3,4};

[0056] Quick branch It includes a first convolutional layer and a first spatial attention module PAM, the feature input is given to the first convolutional layer, and the output of the first convolutional layer is connected to the input of the first spatial attention module PAM;

[0057] The first residual branch b1 includes the first convolutional layer and the second spatial attention module PAM. The feature input is given to the first convolutional layer, and the output of the first convolutional layer is connected to the input of the second spatial attention module PAM.

[0058] The last three residual branches {b i , i=2,3,4} includes the first convolution layer, the second convolution layer, the third convolution layer with a specific expansion rate and the channel attention module CAM, the feature input is given to the first convolution layer, the output of the first convolution layer is connected to the input of the second convolution layer, the output of the second convolution layer is connected to the input of the third convolution layer with a specific expansion rate, and the output of the third convolution layer with a specific expansion rate is connected to the input of the channel attention module CAM;

[0059] The output of the channel attention module CAM is concatenated with the output of the second spatial attention module PAM, and the output after the concatenation is added to the output of the first spatial attention module PAM to output the enhanced features. feature And features

[0060] like Figure 3 As shown in the figure, the background exploration module BEM with integrated attention mechanism is as follows:

[0061] feature Input the background exploration module BEM of the first fusion attention mechanism, features Input the background exploration module BEM of the second fusion attention mechanism, features Input the background exploration module BEM of the third fusion attention mechanism;

[0062] In the third background exploration module BEM fused with attention mechanism, the feature and marginal priors Input to the first background exploration module BA, the first background exploration module BA outputs the updated feature and the updated marginal priors feature and marginal priors Input to the second background exploration module BA, the second background exploration module BA outputs the updated feature and the updated marginal priors feature and marginal priors The input is given to the third background exploration module BA, and the third background exploration module BA outputs the updated edge prior value

[0063] In the background exploration module BEM of the second fusion attention mechanism, the feature and marginal priors Input to the first background exploration module BA, the first background exploration module BA outputs the updated feature and the updated marginal priors feature and marginal priors Input to the second background exploration module BA, the second background exploration module BA outputs the updated feature and the updated marginal priors feature and marginal priors The input is given to the third background exploration module BA, and the third background exploration module BA outputs the updated edge prior value

[0064] In the background exploration module BEM of the first fusion attention mechanism, the feature and marginal priors Input to the first background exploration module BA, the first background exploration module BA outputs the updated feature and the updated marginal priors feature and marginal priors Input to the second background exploration module BA, the second background exploration module BA outputs the updated feature and the updated marginal priors feature and marginal priors The input is given to the third background exploration module BA, and the third background exploration module BA outputs the updated edge prior value

[0065] The background exploration module BA is specifically:

[0066] Marginal Priors After duplication according to the number of channels and features Add and input to the first convolutional layer, which outputs the updated features The output of the first convolutional layer is connected to the input of the second convolutional layer. The output of the second convolutional layer is multiplied by the edge prior value after sigmoid activation and finally multiplied by the edge prior value. Output updated edge priors

[0067] The camouflaged target detection network CABENet, which integrates context perception and background exploration, adopts the following steps when in use:

[0068] Step 1: The disguised target image is input to the Swin-Transformer multi-scale feature extraction module to extract global context information at multiple scales and output feature O i, i = {1, 2, 3, 4}, features O2, O3, and O4 are fused through the feature pyramid to output features O′2, O′3, and O′4;

[0069] Step 2: Features O′2, O′3, and O′4 are input to the attention cascade context perception module AHCM, which uses a parallel multi-branch hybrid convolution mode to expand the receptive field, enhance the global context information features, and use spatial attention and channel attention mechanisms to improve the perception of the target area, and output the enhanced features feature feature

[0070] Step 3: Features feature feature Input to the fully connected decoder module FCD to locate the disguised object, output the output feature C5, and perform a reverse operation on feature C5 to output the edge prior value

[0071] Step 4: Features feature feature The edge prior values ​​are input into the background exploration module BEM of the fusion attention mechanism. The extracted edge prior values ​​are fused with the target features layer by layer using the self-attention mechanism. The edge features of the camouflaged object are extracted by focusing on the background area in the image and the final features are output.

[0072] The above steps are used to detect camouflaged targets.

[0073] Example:

[0074] The present invention conducts experiments in the publicly available camouflaged target detection datasets CAMO, COD10k, and CHAMELEON. CAMO contains 1,250 images, of which 1,000 are used for training and 250 are used for testing. CAMO covers a variety of camouflage forms (such as animal camouflage, camouflage camouflage, and body painting, etc.) in natural and artificial camouflage scenes. CHAMELEON contains 76 manually annotated images. COD10K includes 5,066 camouflaged images, of which 3,040 are used for training and 2,026 are used for testing. It is currently the largest camouflaged object dataset with high-quality pixel-level annotations. This dataset is mainly based on natural camouflage scenes and covers a total of 69 categories of camouflaged objects including flying, aquatic, terrestrial, and amphibians.

[0075] The experimental setting input image size is 352 × 352. During the training process, the epoch is 100, the batch number is 30, and the learning rate is 1e-4.

[0076] In order to compare the proposed model with other advanced methods, the present invention adopts the structural metric S α , average metric E Φ , weighted measure Mean absolute error (MAE) (M) is one of the four widely used indicators to evaluate the performance of COD. α Used to evaluate the structural similarity between the predicted graph and the true value graph; E Φ Used to evaluate the global and local accuracy of the prediction map; A comprehensive measure used to evaluate weighted precision and weighted recall; M is used to evaluate the pixel-level average error between the predicted image and the true image. α 、E Φ , The larger the value of and the smaller the value of M, the better the model performance.

[0077] The present invention provides quantitative evaluation results of CHAMELEON, CAMO, and COD10K datasets. The present invention compares the proposed method with 10 advanced methods in related fields on three benchmark datasets, and the results are shown in Table 1. It can be seen from Table 1 that compared with the advanced COD method, the detection accuracy of the proposed method is significantly better than that of other models. On the CAMO dataset, compared with the SINetV2 method, the proposed method has a better detection accuracy than the SINetV2 method in terms of structural metric S. α , average enhancement measure E Φ , weighted F-measure The average absolute error was improved by 5.6%, 5.6%, and 8.9% respectively. M A decrease of 2.9%.

[0078] Table 1 Quantitative comparison

[0079]

[0080] In terms of visual inspection, the comparison results of the proposed model with other 6 representative models are as follows: Figure 4 As shown. Figure 4It can be seen from the figure that the proposed method can segment the camouflaged object more accurately and completely. For example, it can be seen from the first and fifth rows that for objects with low contrast and complex boundary structures, the camouflaged object image obtained by the proposed method is more obvious and clearer than that obtained by other methods. This is due to the background exploration module that integrates the attention mechanism. By exploring the background area to enhance the boundary features, the boundary of the target is clearer and more specific. It can be seen from the second and third rows that the proposed method can locate camouflaged targets with different scales. This is due to the attention cascade context perception module. By effectively integrating feature information at different levels, the model can be more flexible and accurate when processing multi-scale targets. It can be seen from the fourth row that the proposed method can detect multiple camouflaged objects at the same time. This is due to the global capture of image features by the swin-transformer network. It can be seen from the last row that the target is partially obscured by the surrounding environment, but the proposed model can still successfully detect the camouflaged target.

[0081] In order to verify the effectiveness of each module in the model of the present invention, ablation experiments were carried out on each submodule. The quantitative results are shown in Table 2 and the visual results are shown in Figure 5 As shown in the figure. Method 1 shows that the swin-transformer and feature fusion parts are used as the baseline model (Baseline) for experiments. Methods 2 and 3 show that the attention cascade context perception module (AHCM) and the background exploration module (BEM) with fusion attention mechanism are added to the baseline for experiments. Method 4 shows the experimental results of the overall model. From Method 2, it can be seen that compared with the single basic network, Baseline+AHCM has a better performance on the COD10k dataset. α 、E Φ , They increased by 0.3%, 0.6%, and 1.2%, respectively. M It fell by 0.2% through Figure 5 Comparing the third and fourth columns, we can see that the camouflaged target can be detected more clearly, which further proves from the visual effect that AHCM is necessary for improving the model performance. From method 3, we can see that compared with the single basic network, Baseline+BEM has a better performance on the COD10k dataset. α 、E Φ , They increased by 0.9%, 1.4%, and 3.1% respectively, and M decreased by 0.3%. Figure 5 Comparing the third and fifth columns, we can see that the BEM module makes the boundary structure of the disguised target clearer. From method 4, we can see that the performance of combining multiple modules is improved compared with using a single module.

[0082] Table 2 Ablation experiment

[0083]

[0084] From the above, it can be seen that the present invention proposes a camouflaged target detection method that integrates context perception and background exploration to address the technical problems of incomplete results and blurred edge details in COD. This method uses swin-transformer to achieve multi-scale feature representation. In the attention cascade context perception module, a dual attention mechanism is used to help the model perform more fine-grained feature interaction and integration between different spatial scales and channels. Finally, a background exploration module that integrates the attention mechanism is used to refine the rough prediction from the deep layer. The experimental results on three public datasets show that the method of the present invention achieves better detection results than other methods. In future work, new network models will be further explored to enable the COD method to achieve more accurate camouflaged target detection.

Claims

1. A camouflaged target detection network integrating context perception and background exploration, characterized in that: The network includes a multi-scale feature extraction module Swin-Transformer, an attention cascade context perception module AHCM, a fully connected decoder module FCD, and a background exploration module BEM that integrates the attention mechanism; The output of the multi-scale feature extraction module Swin-Transformer is connected to the input of the attention cascade context perception module AHCM, the output of the attention cascade context perception module AHCM is connected to the fully connected decoder module FCD and the input of the background exploration module BEM fused with the attention mechanism, the output of the fully connected decoder module FCD is connected to the input of the background exploration module BEM fused with the attention mechanism, and the output of the background exploration module BEM fused with the attention mechanism is used to output the final prediction.

2. The network according to claim 1, characterized in that The multi-scale feature extraction module Swin-Transformer is used as the feature extraction layer to output multi-scale features O i ,i={1,2,3,4}, features O2, O3, and O4 are fused through the feature pyramid to output features O2′, O3′, and O4′; The feature O2′ is input to the first attention cascade context perception module AHCM, and the output of the first attention cascade context perception module AHCM is connected to the input of the fully connected decoder module FCD and the input of the background exploration module BEM of the first fusion attention mechanism; The feature O3′ is input to the second attention cascade context perception module AHCM, and the output of the second attention cascade context perception module AHCM is connected to the input of the fully connected decoder module FCD and the input of the background exploration module BEM of the second fusion attention mechanism; The feature O4′ is input to the third attention cascade context perception module AHCM, and the output of the third attention cascade context perception module AHCM is connected to the input of the fully connected decoder module FCD and the input of the background exploration module BEM of the third fusion attention mechanism; The fully connected decoder module FCD outputs feature C5, which performs a reverse operation to output the edge prior value. Marginal Priors The input is given to the background exploration module BEM of the third fusion attention mechanism, and the output of the background exploration module BEM of the third fusion attention mechanism is pixel-added with the feature C5 to output the feature C4; Feature C4 is input to the background exploration module BEM of the second fusion attention mechanism, and the output of the background exploration module BEM of the second fusion attention mechanism is pixel-added with feature C4 to output feature C3; Feature C3 is input to the background exploration module BEM of the first fusion attention mechanism, and the output of the background exploration module BEM of the first fusion attention mechanism is pixel-added with feature C3 to output the final feature.

3. The network according to claim 1 or 2, characterized in that: The attention cascade context perception module AHCM is specifically: Any input of feature O2′, feature O3′, feature O4′ is given to the shortcut branch 4 parallel residual branches {b i ,i=1,2,3,4}; Quick branch It includes a first convolutional layer and a first spatial attention module PAM, the feature input is given to the first convolutional layer, and the output of the first convolutional layer is connected to the input of the first spatial attention module PAM; The first residual branch b1 includes the first convolutional layer and the second spatial attention module PAM. The feature input is given to the first convolutional layer, and the output of the first convolutional layer is connected to the input of the second spatial attention module PAM. The last three residual branches {b i , i=2,3,4} includes the first convolution layer, the second convolution layer, the third convolution layer with a specific expansion rate and the channel attention module CAM, the feature input is given to the first convolution layer, the output of the first convolution layer is connected to the input of the second convolution layer, the output of the second convolution layer is connected to the input of the third convolution layer with a specific expansion rate, and the output of the third convolution layer with a specific expansion rate is connected to the input of the channel attention module CAM; The output of the channel attention module CAM is concatenated with the output of the second spatial attention module PAM, and the output after the concatenation is added to the output of the first spatial attention module PAM to output the enhanced features. feature And features 4. The network according to claim 1 or 2, characterized in that: The background exploration module BEM that integrates the attention mechanism is as follows: feature Input the background exploration module BEM of the first fusion attention mechanism, features Input the background exploration module BEM of the second fusion attention mechanism, features Input the background exploration module BEM of the third fusion attention mechanism; In the third background exploration module BEM fused with attention mechanism, the feature and marginal priors Input to the first background exploration module BA, the first background exploration module BA outputs the updated feature and the updated marginal priors feature and marginal priors Input to the second background exploration module BA, the second background exploration module BA outputs the updated feature and the updated marginal priors feature and marginal priors The input is given to the third background exploration module BA, and the third background exploration module BA outputs the updated edge prior value In the background exploration module BEM of the second fusion attention mechanism, the feature and marginal priors Input to the first background exploration module BA, the first background exploration module BA outputs the updated feature and the updated marginal priors feature and marginal priors Input to the second background exploration module BA, the second background exploration module BA outputs the updated feature and the updated marginal priors feature and marginal priors The input is given to the third background exploration module BA, and the third background exploration module BA outputs the updated edge prior value In the background exploration module BEM of the first fusion attention mechanism, the feature and marginal priors Input to the first background exploration module BA, the first background exploration module BA outputs the updated feature and the updated marginal priors feature and marginal priors Input to the second background exploration module BA, the second background exploration module BA outputs the updated feature and the updated marginal priors feature and marginal priors The input is given to the third background exploration module BA, and the third background exploration module BA outputs the updated edge prior value The background exploration module BA is specifically: Marginal Priors After duplication according to the number of channels and features Add and input to the first convolutional layer, which outputs the updated features The output of the first convolutional layer is connected to the input of the second convolutional layer. The output of the second convolutional layer is multiplied by the edge prior value after sigmoid activation and finally multiplied by the edge prior value. Output updated edge priors 5. The network according to claim 1, when in use, adopts the following steps: Step 1: The disguised target image is input to the Swin-Transformer multi-scale feature extraction module to extract global context information at multiple scales and output feature O i ,i={1,2,3,4}, features O2, O3, and O4 are fused through the feature pyramid to output features O2′, O3′, and O4′; Step 2: Features O2′, O3′, and O4′ are input to the attention cascade context perception module AHCM, which uses a parallel multi-branch hybrid convolution mode to expand the receptive field, enhance the global context information features, and use spatial attention and channel attention mechanisms to improve the perception of the target area, and output the enhanced features feature feature Step 3: Features feature feature Input to the fully connected decoder module FCD to locate the disguised object, output the output feature C5, and perform a reverse operation on feature C5 to output the edge prior value Step 4: Features feature feature The edge prior values ​​are input into the background exploration module BEM of the fusion attention mechanism. The extracted edge prior values ​​are fused with the target features layer by layer using the self-attention mechanism. The edge features of the camouflaged object are extracted by focusing on the background area in the image and the final features are output. The above steps are used to detect camouflaged targets.