Hierarchical attention and feature refinement fused camouflage target detection network and method
By integrating layered attention and feature refinement, the problems of incomplete detection results and excessive calculation costs in the existing technology are solved, and precise positioning and fine segmentation of camouflage objects are achieved, and the calculation costs are reduced. And improve detection performance.
Patent Information
- Application Number
- CN202411477555.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-22
- Publication Date
- 2025-02-18
AI Technical Summary
The existing camouflage object detection methods have problems such as incomplete detection results and excessive calculation costs, and it is difficult to achieve precise positioning and fine segmentation of camouflage objects.
A camouflage object detection network that integrates layered attention and feature refinement is adopted. This network includes a multi-scale feature extraction module, a layered attention network, a fully connected decoder module and feature refinement network, through multi-level feature interaction and refinement , realize accurate positioning and fine segmentation of camouflage objects.
The precise positioning and fine segmentation of camouflage objects is achieved, which reduces computational costs and improves detection performance, and can accurately detect camouflage objects in scenarios with low contrast and complex boundary structures.
Smart Images
Figure CN119478344A_ABST
Abstract
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 of a disguised target detection network integrating hierarchical attention and feature refinement. Background Art
[0002] Camouflage is a common defense strategy in nature. Some species use their appearance similar to the surrounding environment to hide themselves and avoid being discovered by predators. Camouflaged object detection (COD) aims to detect camouflaged targets 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 manually designed low-level features, such as direct visual features such as color, texture, and optical flow. In complex scenes with extremely low contrast between foreground and background, problems such as time-consuming manual feature extraction, insufficient transferability, and poor detection performance are often encountered. To address these problems, researchers have proposed a variety of camouflaged 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 remove the interference of irrelevant features. Subsequently, they proposed the SINetV2 model, which mines the edge information of the camouflaged target through a group-guided inverted attention module to generate more refined predictions. However, in the group-guided process, the distribution of feature representations may be uneven. Mei et al. constructed a COD network PFNet consisting of a localization module and a focusing module. The network first locates potential targets from a global perspective, and then gradually predicts camouflaged targets by focusing on the blurred area. Then the camouflaged target is gradually predicted by focusing on the blurred area. However, in the focusing process, directly multiplying the current feature with the prediction prior may cause feature confusion. Zhang et al. proposed an attention-guided semantic and boundary interaction network, which uses contrastive learning to separate the camouflaged object from the background, thereby accurately obtaining the location of the camouflaged target. Although the above methods have made some progress in the COD field, most COD methods find it difficult to achieve accurate target segmentation because the camouflaged object is highly similar to the background and is blurred at the edge, and usually require a larger number of parameters and higher computational complexity to pursue better performance. Summary of the invention
[0004] The purpose of the present invention is to solve the problems of incomplete detection results and high computational cost of existing camouflaged target detection methods at the same time, and to realize accurate positioning and fine segmentation of camouflaged objects by effectively aggregating and utilizing semantic and detail information, and to propose a camouflaged target detection network and method that integrates hierarchical attention and feature refinement.
[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is:
[0006] A camouflaged target detection network that integrates hierarchical attention and feature refinement, which includes a feature extraction layer SMT, a hierarchical attention network layer TAN, a fully connected decoder module FCD, and a feature refinement network layer FRN;
[0007] The output of the feature extraction layer SMT is connected to the input of the hierarchical attention network layer TAN, the output of the hierarchical attention network layer TAN is connected to the fully connected decoder module FCD and the input of the feature refinement network layer FRN, the output of the fully connected decoder module FCD is connected to the input of the feature refinement network layer FRN, and the output of the feature refinement network layer FRN is used to output the final prediction.
[0008] The multi-scale feature extraction module SMT is used as the feature extraction layer to output the multi-scale feature f i ,i={1,2,3,4}.
[0009] The feature f1 is input to the fourth attention-based compression activation module ASEM in the hierarchical attention network layer TAN, and the output end of the fourth attention-based compression activation module ASEM outputs the feature F1. The feature F1 is input to the third attention-based compression activation module ASEM, the fully connected decoder module FCD, and the fourth enhanced convolutional block attention module ECAM in the feature refinement network layer FRN;
[0010] Feature f2 is input to the third attention-based compression activation module ASEM in the hierarchical attention network layer TAN. The output end of the third attention-based compression activation module ASEM outputs feature F2. Feature F2 is input to the second attention-based compression activation module ASEM and the fully connected decoder module FCD. Feature F2 and F1 are added together to output feature F2′. Feature F2′ is input to the third enhanced convolutional block attention module ECAM in the feature refinement network layer FRN.
[0011] Feature f3 is input to the second attention-based compression activation module ASEM in the hierarchical attention network layer TAN. The output end of the second attention-based compression activation module ASEM outputs feature F3. Feature F3 is input to the first attention-based compression activation module ASEM and the fully connected decoder module FCD. Feature F3, feature F2 and feature F1 are added together to output feature F3′. Feature F3′ is input to the second enhanced convolutional block attention module ECAM in the feature refinement network layer FRN.
[0012] Feature f4 is input to the first attention-based compression activation module ASEM in the hierarchical attention network layer TAN. The output end of the first attention-based compression activation module ASEM outputs feature F4. Feature F4, feature F3, feature F2 and feature F1 are added together to output feature F4′. Feature F4′ is input to the first enhanced convolutional block attention module ECAM in the feature refinement network layer FRN.
[0013] The output feature of the fully connected decoder module FCD is feature C, and feature C is pixel-added with the output of the fourth enhanced convolutional block attention module ECAM in the feature refinement network layer FRN to output feature C1;
[0014] Feature C1 is input to the third enhanced convolutional block attention module ECAM in the feature refinement network layer FRN, and the output of the third enhanced convolutional block attention module ECAM is pixel-added with feature C to output feature C2;
[0015] Feature C2 is input to the second enhanced convolutional block attention module ECAM in the feature refinement network layer FRN, and the output of the second enhanced convolutional block attention module ECAM is pixel-added with feature C to output feature C3;
[0016] Feature C3 is input to the first enhanced convolutional block attention module ECAM in the feature refinement network layer FRN, and the first enhanced convolutional block attention module ECAM outputs the final prediction;
[0017] The attention-based compression activation module ASEM is specifically:
[0018] Feature i and feature F i-1 Input to the first split layer and output features f h ′ and feature f l ′, feature f h ′ and feature f l ′After connection, it is input to the convolution layer, and the convolution layer outputs the feature F in , feature F in Input to two one-dimensional average pooling layers and output features P respectively h ′ and feature P l ′, feature P h′ and feature P l ′ is input to the ConvBS layer, and the ConvBS layer outputs the feature P c ′, feature P c 'Input to the second split layer output feature X w and feature X h , feature X w and feature X h Input to the coordinate attention layer output coordinate attention map, coordinate attention map and feature F in The element-wise multiplication outputs feature X, which is input to the average pooling layer. The output of the average pooling layer is connected to the input of the first FC layer, the output of the first FC layer is connected to the input of the ReLU layer, the output of the ReLU layer is connected to the input of the second FC layer, the output of the second FC layer is connected to the input of the sigmoid layer, the output of the sigmoid layer is multiplied with feature X to output feature X′, feature X′ and feature F in After multiplication, the output feature F i .
[0019] The enhanced convolutional block attention module ECAM is specifically:
[0020] Feature C i-1 and feature F i ′ is connected and input to the first point-to-point layer, the first point-to-point layer outputs feature F, feature F is input to the split layer to output feature F′ and feature F″, feature F′ is input to the convolutional layer, the output of the convolutional layer is connected with feature F″ and input to the second point-to-point layer, the output feature h of the second point-to-point layer, feature h is input to the channel attention module, the output of the channel attention module is multiplied by feature h to output feature feature h′, feature h′ is input to the spatial attention module, the output of the spatial attention module is multiplied by feature h′ and input to the third point-to-point layer, the output of the third point-to-point layer is added with feature F and up-sampled to output the final feature.
[0021] The disguised object detection network that integrates hierarchical attention and feature refinement takes the following steps when used:
[0022] Step 1: The disguised target image is input to the feature extraction layer SMT to extract global context information at multiple scales and output features f i ,i={1,2,3,4}.
[0023] Step 2: Feature f i , i = {1, 2, 3, 4} is input to the hierarchical attention network layer TAN, which uses a top-down approach to promote the interaction between multi-scale features, uses four compression activation modules ASEM to perceive features of different dimensions, and outputs feature F i ,i={1,2,3,4}.
[0024] Step 3: Features F1, F2, and F3 are input to the fully connected decoder module FCD to locate the disguised object and output the output feature C.
[0025] Step 4: Features F1, F2, and F3 are input to the feature refinement network layer FRN, which uses multi-level features to guide low-level features. Four enhanced convolutional block attention modules ECAM are used to further refine the context information. A top-down approach is used to aggregate multi-scale features to achieve the goal of refined prediction and output the final prediction.
[0026] The above steps are used to detect camouflaged targets.
[0027] Compared with the prior art, the present invention has the following technical effects:
[0028] 1) In order to simultaneously solve the problems of incomplete detection results and high computational cost of disguised target detection methods, the present invention proposes a disguised target detection network and method that integrates hierarchical attention and feature refinement, which achieves accurate positioning and fine segmentation of disguised objects through effective aggregation and utilization of semantic and detail information.
[0029] 2) The model of the present invention uses SMT as the network skeleton to extract image features, effectively integrates and perceives multi-scale information through the hierarchical attention network layer TAN, uses FCD to effectively locate hidden objects, and finally the feature refinement network layer FRN further refines the camouflage prediction results.
[0030] 3) This paper proposes a hierarchical attention network layer TAN, which uses four attention-based compressed activation modules ASEM to perceive features of different dimensions, utilizes a top-down approach to promote the interaction between multi-scale features, and effectively propagates high-level semantic information to low-level features to better handle disguised objects of different sizes.
[0031] 4) The present invention proposes a feature refinement network layer FRN, which further refines the contextual information by using four enhanced convolutional block attention modules ECAM. This network layer adopts a top-down approach to aggregate multi-scale features, thereby retaining more detailed information and achieving the goal of refined prediction. BRIEF DESCRIPTION OF THE DRAWINGS
[0032] The present invention will be further described below in conjunction with the accompanying drawings and embodiments:
[0033] Figure 1 The overall framework diagram of the camouflaged target detection network and method that integrates hierarchical attention and feature refinement in the present invention;
[0034] Figure 2 for Figure 1 Schematic diagram of the attention-based compressed activation module ASEM;
[0035] Figure 3 for Figure 1 Schematic diagram of the enhanced convolutional block attention module ECAM;
[0036] Figure 4 is a comparison diagram of the model of the present invention and 6 other representative models;
[0037] Figure 5 This is a visualization result diagram of the ablation experiment in the present invention. DETAILED DESCRIPTION
[0038] like Figure 1 As shown, a camouflaged target detection network and method integrating hierarchical attention and feature refinement, the network includes a feature extraction layer SMT, a hierarchical attention network layer TAN, a fully connected decoder module FCD and a feature refinement network layer FRN;
[0039] The output of the feature extraction layer SMT is connected to the input of the hierarchical attention network layer TAN, the output of the hierarchical attention network layer TAN is connected to the fully connected decoder module FCD and the input of the feature refinement network layer FRN, the output of the fully connected decoder module FCD is connected to the input of the feature refinement network layer FRN, and the output of the feature refinement network layer FRN is used to output the final prediction.
[0040] The multi-scale feature extraction module SMT is used as the feature extraction layer to output the multi-scale feature f i ,i={1,2,3,4}.
[0041] The feature f1 is input to the fourth attention-based compression activation module ASEM in the hierarchical attention network layer TAN, the output feature F1 of the fourth attention-based compression activation module ASEM, and the feature F1 is input to the third attention-based compression activation module ASEM, the fully connected decoder module FCD, and the fourth enhanced convolutional block attention module ECAM in the feature refinement network layer FRN;
[0042] Feature f2 is input to the third attention-based compression activation module ASEM in the hierarchical attention network layer TAN. The output feature F2 of the third attention-based compression activation module ASEM is input to the second attention-based compression activation module ASEM and the fully connected decoder module FCD. After adding the pixels of features F2 and F1, feature F2′ is output. Feature F2′ is input to the third enhanced convolutional block attention module ECAM in the feature refinement network layer FRN.
[0043] Feature f3 is input to the second attention-based compression activation module ASEM in the hierarchical attention network layer TAN. The output feature F3 of the second attention-based compression activation module ASEM is input to the first attention-based compression activation module ASEM and the fully connected decoder module FCD. After adding the pixels of feature F3, feature F2 and feature F1, feature F3′ is output. Feature F3′ is input to the second enhanced convolutional block attention module ECAM in the feature refinement network layer FRN.
[0044] Feature f4 is input to the first attention-based compression activation module ASEM in the hierarchical attention network layer TAN. The output feature F4 of the first attention-based compression activation module ASEM is output. Feature F4, feature F3, feature F2 and feature F1 are added together to output feature F4′. Feature F4′ is input to the first enhanced convolutional block attention module ECAM in the feature refinement network layer FRN.
[0045] The fully connected decoder module FCD outputs the feature C, which is pixel-wise added to the output of the fourth enhanced convolutional block attention module ECAM in the feature refinement network layer FRN to output feature C1;
[0046] Feature C1 is input to the third enhanced convolutional block attention module ECAM in the feature refinement network layer FRN, and the output of the third enhanced convolutional block attention module ECAM is pixel-added with feature C to output feature C2;
[0047] Feature C2 is input to the second enhanced convolutional block attention module ECAM in the feature refinement network layer FRN, and the output of the second enhanced convolutional block attention module ECAM is pixel-added with feature C to output feature C3;
[0048] Feature C3 is input to the first enhanced convolutional block attention module ECAM in the feature refinement network layer FRN, and the first enhanced convolutional block attention module ECAM outputs the final prediction;
[0049] like Figure 2 As shown in Figure 1, the attention-based compression activation module ASEM is specifically:
[0050] Feature i and feature F i-1 Input to the first split layer and output features f h ′ and feature f l ′, feature f h ′ and feature f l ′After connection, it is input to the convolution layer, and the convolution layer outputs the feature F in , feature F in Input to two one-dimensional average pooling layers and output features P respectively h ′ and feature P l ′, feature Ph ′ and feature P l ′ is input to the ConvBS layer, and the ConvBS layer outputs the feature P c ′, feature P c 'Input to the second split layer output feature X w and feature X h , feature X w and feature X h Input to the coordinate attention layer output coordinate attention map, coordinate attention map and feature F in The element-wise multiplication outputs feature X, which is input to the average pooling layer. The output of the average pooling layer is connected to the input of the first FC layer, the output of the first FC layer is connected to the input of the ReLU layer, the output of the ReLU layer is connected to the input of the second FC layer, the output of the second FC layer is connected to the input of the sigmoid layer, the output of the sigmoid layer is multiplied with feature X to output feature X′, feature X′ and feature F in After multiplication, the output feature F i .
[0051] like Figure 3 As shown, the enhanced convolutional block attention module ECAM is specifically:
[0052] Feature C i-1 and feature F i ′ is connected and input to the first point-to-point layer, the first point-to-point layer outputs feature F, feature F is input to the split layer to output feature F′ and feature F″, feature F′ is input to the convolutional layer, the output of the convolutional layer is connected with feature F″ and input to the second point-to-point layer, the output feature h of the second point-to-point layer, feature h is input to the channel attention module, the output of the channel attention module is multiplied by feature h to output feature feature h′, feature h′ is input to the spatial attention module, the output of the spatial attention module is multiplied by feature h′ and input to the third point-to-point layer, the output of the third point-to-point layer is added with feature F and up-sampled to output the final feature.
[0053] When using the camouflaged target detection network and method HARNet that integrates hierarchical attention and feature refinement, the following steps are taken:
[0054] Step 1: The disguised target image is input to the feature extraction layer SMT to extract global context information at multiple scales and output features f i ,i={1,2,3,4}.
[0055] Step 2: Feature f i , i = {1, 2, 3, 4} is input to the hierarchical attention network layer TAN, which uses a top-down approach to promote the interaction between multi-scale features, uses four compression activation modules ASEM to perceive features of different dimensions, and outputs feature F i,i={1,2,3,4}.
[0056] Step 3: Features F1, F2, and F3 are input to the fully connected decoder module FCD to locate the disguised object and output the output feature C.
[0057] Step 4: Features F1, F2, and F3 are input to the feature refinement network layer FRN, which uses multi-level features to guide low-level features. Four enhanced convolutional block attention modules ECAM are used to further refine the context information. A top-down approach is used to aggregate multi-scale features to achieve the goal of refined prediction and output the final prediction.
[0058] The above steps are used to detect camouflaged targets.
[0059] Example:
[0060] 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.
[0061] The experimental setting input image size is 384 × 384. During the training process, the epoch is 100, the batch number is 10, and the learning rate is 5e-4.
[0062] 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. In addition, the present invention also uses parameters (Params), computational complexity (FLOPs) and frames per second (FPS) to evaluate the computational efficiency of the model.
[0063] The present invention provides quantitative evaluation results of CHAMELEON, CAMO, and COD10K datasets and the computational efficiency of the model. The present invention compares the proposed method with 9 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 COD method based on CNN backbone, the detection accuracy of the method of the present invention is significantly better than that of other models. Compared with the COD method based on transformer backbone, the present invention achieves better performance with the lowest parameter amount (14.04M) and computational cost (11.39G).
[0064] Table 1 Quantitative comparison
[0065]
[0066] 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 4 It 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 other methods. This is due to the feature refinement network layer FRN, which enhances the boundary features by locally refining features to make the boundary of the target 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, which is due to the hierarchical attention network layer TAN, which effectively integrates feature information at different levels, making the model 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, which is due to the global capture of image features by the SMT 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.
[0067] 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 5As shown in Figure 2, Method 1 uses the feature extraction layer SMT as the baseline model (Baseline) for experiments. Methods 2 and 3 respectively add the attention-based compressed activation module ASEM and the enhanced convolutional block attention module ECAM 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+ASEM has a better performance on the COD10k dataset. α 、E Φ , They increased by 1.2%, 0.9%, and 0.8% respectively. 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 ASEM is necessary for improving the model performance. From method 3, we can see that compared with the single basic network, Baseline+ECAM has a better performance on the CAMO dataset. α 、E Φ , They increased by 1.3%, 1.1%, and 1.1% respectively, and M decreased by 0.2%. Figure 5 Comparison between columns 3 and 5 shows that the ECAM module makes the boundary structure of the camouflaged target clearer. From method 4, it can be seen that the performance of combining multiple modules is improved compared with using a single module.
[0068]
[0069] As can be seen from the above, in order to solve the problems of incomplete detection results and high computational cost of camouflaged target detection methods, the present invention proposes a camouflaged target detection network and method that integrates hierarchical attention and feature refinement. The method uses the SMT network to achieve multi-scale feature representation. In the hierarchical attention network layer TAN, the compressed activation module ASEM helps the model to perform finer-grained feature interaction and integration between different spatial scales and channels. Finally, the feature refinement network layer FRN is used to make rough predictions from deep layers. 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 object detection network that integrates hierarchical attention and feature refinement, characterized by: The network includes a feature extraction layer SMT, a layered attention network layer TAN, a fully connected decoder module FCD, and a feature refinement network layer FRN; The output of the feature extraction layer SMT is connected to the input of the hierarchical attention network layer TAN, the output of the hierarchical attention network layer TAN is connected to the fully connected decoder module FCD and the input of the feature refinement network layer FRN, the output of the fully connected decoder module FCD is connected to the input of the feature refinement network layer FRN, and the output of the feature refinement network layer FRN is used to output the final prediction.
2. The network according to claim 1, characterized in that The multi-scale feature extraction module SMT is used as the feature extraction layer to output the multi-scale feature f i ,i={1,2,3,4}; The feature f1 is input to the fourth attention-based compression activation module ASEM in the hierarchical attention network layer TAN, and the output end of the fourth attention-based compression activation module ASEM outputs the feature F1. The feature F1 is input to the third attention-based compression activation module ASEM, the fully connected decoder module FCD, and the fourth enhanced convolutional block attention module ECAM in the feature refinement network layer FRN; Feature f2 is input to the third attention-based compression activation module ASEM in the hierarchical attention network layer TAN. The output end of the third attention-based compression activation module ASEM outputs feature F2. Feature F2 is input to the second attention-based compression activation module ASEM and the fully connected decoder module FCD. Feature F2 and F1 are added together to output feature F2′. Feature F2′ is input to the third enhanced convolutional block attention module ECAM in the feature refinement network layer FRN. Feature f3 is input to the second attention-based compression activation module ASEM in the hierarchical attention network layer TAN. The output end of the second attention-based compression activation module ASEM outputs feature F3. Feature F3 is input to the first attention-based compression activation module ASEM and the fully connected decoder module FCD. Feature F3, feature F2 and feature F1 are added together to output feature F3′. Feature F3′ is input to the second enhanced convolutional block attention module ECAM in the feature refinement network layer FRN. Feature f4 is input to the first attention-based compression activation module ASEM in the hierarchical attention network layer TAN. The output end of the first attention-based compression activation module ASEM outputs feature F4. Feature F4, feature F3, feature F2 and feature F1 are added together to output feature F4′. Feature F4′ is input to the first enhanced convolutional block attention module ECAM in the feature refinement network layer FRN. The output feature of the fully connected decoder module FCD is feature C, and feature C is pixel-added with the output of the fourth enhanced convolutional block attention module ECAM in the feature refinement network layer FRN to output feature C1; Feature C1 is input to the third enhanced convolutional block attention module ECAM in the feature refinement network layer FRN, and the output of the third enhanced convolutional block attention module ECAM is pixel-added with feature C to output feature C2; Feature C2 is input to the second enhanced convolutional block attention module ECAM in the feature refinement network layer FRN, and the output of the second enhanced convolutional block attention module ECAM is pixel-added with feature C to output feature C3; Feature C3 is input to the first enhanced convolutional block attention module ECAM in the feature refinement network layer FRN, and the first enhanced convolutional block attention module ECAM outputs the final prediction.
3. The network according to claim 1 or 2, characterized in that: The attention-based compression activation module ASEM is specifically: Feature i and feature F i-1 Input to the first split layer and output features f h ′ and feature f l ′, feature f h ′ and feature f l ′After connection, it is input to the convolution layer, and the convolution layer outputs the feature F in , feature F in Input to two one-dimensional average pooling layers and output features P respectively h ′ and feature P l ′, feature P h ′ and feature P l ′ is input to the ConvBS layer, and the ConvBS layer outputs the feature P c ′, feature P c 'Input to the second split layer output feature X w and feature X h , feature X w and feature X h Input to the coordinate attention layer output coordinate attention map, coordinate attention map and feature F in The element-wise multiplication outputs feature X, which is input to the average pooling layer. The output of the average pooling layer is connected to the input of the first FC layer, the output of the first FC layer is connected to the input of the ReLU layer, the output of the ReLU layer is connected to the input of the second FC layer, the output of the second FC layer is connected to the input of the sigmoid layer, the output of the sigmoid layer is multiplied with feature X to output feature X′, feature X′ and feature F in After multiplication, the output feature F i .
4. The network according to claim 1 or 2, characterized in that: The enhanced convolutional block attention module ECAM is specifically: Feature C i-1 and feature F i ′ is connected and input to the first point-to-point layer, the first point-to-point layer outputs feature F, feature F is input to the split layer to output feature F′ and feature F″, feature F′ is input to the convolutional layer, the output of the convolutional layer is connected with feature F″ and input to the second point-to-point layer, the output feature h of the second point-to-point layer, feature h is input to the channel attention module, the output of the channel attention module is multiplied by feature h to output feature feature h′, feature h′ is input to the spatial attention module, the output of the spatial attention module is multiplied by feature h′ and input to the third point-to-point layer, the output of the third point-to-point layer is added with feature F and up-sampled to output the final feature.
5. The network according to claim 1, characterized in that The camouflaged target detection network integrating hierarchical attention and feature refinement adopts the following steps when used: Step 1: The disguised target image is input to the feature extraction layer SMT to extract global context information at multiple scales and output features f i ,i={1,2,3,4}; Step 2: Feature f i , i = {1, 2, 3, 4} is input to the hierarchical attention network layer TAN, which uses a top-down approach to promote the interaction between multi-scale features, uses four compression activation modules ASEM to perceive features of different dimensions, and outputs feature F i ,i={1,2,3,4}; Step 3: Features F1, F2, and F3 are input to the fully connected decoder module FCD to locate the disguised object and output the output feature C; Step 4: Features F1, F2, and F3 are input to the feature refinement network layer FRN, which uses multi-level features to guide low-level features. Four enhanced convolutional block attention modules ECAM are used to further refine the context information. A top-down approach is used to aggregate multi-scale features to achieve the goal of refined prediction and output the final prediction. The above steps are used to detect camouflaged targets.