Pixshuffle feature amplification and multi-feature fusion-based camouflage target detection method

Abstract

The invention discloses a camouflage target detection method based on Pixshuffle feature amplification and multi-feature fusion, the method is used for camouflage target detection PAFNet network detection, a network comprises a searching and positioning module, an amplification and fusion module and a feature refining module, the searching and positioning module is composed of a backbone network and a cavity space convolution pooling pyramid improved version, and the amplification and fusion module is composed of an amplification module and a feature refining module. The amplification and fusion module is composed of a CBR module, three PSFF modules and a Pixshuffle module, and the feature refining module is composed of three ERM modules and an ORM module. In the feature fusion stage, a super-resolution up-sampling method is used for feature amplification, multi-feature fusion is carried out through a PSFF module, two adjacent features and forward transfer features are fused, the huge difference between deep semantic features and shallow detail features is solved, in the feature refinement stage, an expansion algorithm is used for paying attention to targets and edges, and the accuracy of feature refinement is improved. Therefore, highly similar foreground and background can be accurately separated.

Description

Technical Field

[0001] The invention belongs to the technical field of computer vision, and in particular relates to a camouflaged target detection method based on Pixshuffle feature amplification and multi-feature fusion. Background Art

[0002] Disguised target detection is a challenging task to detect disguised targets from images. It is necessary to fight against excellent camouflage strategies and distinguish the subtle differences between foreground objects and backgrounds. With the development of target detection technology, disguised target detection technology based on deep networks makes it possible to detect disguised targets. Compared with human eye detection, disguised target detection technology can quickly and accurately detect disguised targets, improve detection efficiency and accuracy, and save labor costs. Disguised target detection technology has a wide range of application value in many fields.

[0003] At present, camouflaged target detection mainly faces two major challenges: one is that the foreground object is highly similar to the background color and texture, making it difficult to locate the object; the other is the blurred edge. Even if the object is accurately located, it is difficult to accurately segment the object due to its ambiguous edge information. In recent years, in order to solve these problems, SINet and PFNet divide the camouflaged target detection process into two stages: search and recognition. The search stage roughly locates the camouflaged target, and the recognition stage refines the rough map to obtain an accurate prediction map. Due to the high similarity between the foreground object and the background color, the rough positioning in the search stage is inaccurate, which in turn leads to inaccurate segmentation results. The FEDER network uses ASPP (atrous spatial convolutional pooling pyramid) to locate the deepest features in the search stage, and the positioning effect is improved, but there is still an inaccurate problem for the positioning of small targets; SARNet noticed that the size of the resolution is crucial to the detection of camouflaged targets, so it uses bicubic interpolation to upsample the features and then perform feature fusion. Although interpolation upsampling can increase the resolution of an image, its ability to actually restore details in the image is poor. The added pixels are usually calculated based on existing pixels and cannot restore details that were not captured in the original image. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a camouflaged target detection method based on Pixshuffle feature amplification and multi-feature fusion, which aims to solve the problems of inaccurate positioning of small targets and weak detail information restoration by upsampling.

[0005] In order to solve the above technical problems, the present invention is implemented in the following ways:

[0006] A camouflaged target detection method based on Pixshuffle feature amplification and multi-feature fusion is used for camouflaged target detection PAFNet network detection. The network includes a search and positioning module (SPM), a magnification and fusion module (MFM) and a feature refinement module (FRM). The search and positioning module is composed of a backbone network PVTv2 and an improved version of a dilated spatial convolutional pooling pyramid (SE-ASPP). The magnification and fusion module is composed of a CBR module, three Pixshuffle Feature Fusion (PSFF) modules and a Pixshuffle module. The feature refinement module is composed of three Edge Refinement Module (ERM) modules, an Object Refinement Module (ORM) module and a CBR module.

[0007] Furthermore, the backbone network in the search and positioning module uses PVTv2-B3 to extract features of different scales, and the input image I∈R 3×H×W , and extract features of different scales through four Block modules in turn Then the deepest level feature t4 is input into the SE-ASPP module for coarse positioning to obtain the initial prediction map p5.

[0008] Preferably, the specific method of obtaining the initial prediction map p5 by coarse positioning is as follows:

[0009] The deepest level feature t4 is input into five different branches in the SE-ASPP module, which includes three 3×3 hole convolution layers with expansion rates of 6, 12 and 18, a pooling layer and a 1×1 convolution layer. The features obtained from the five different branches are respectively input into the SE channel attention module for feature extraction. Finally, the enhanced features of the five branches are fused, and then the channel is reduced through the 1×1 convolution layer to obtain the coarse positioning initial prediction map p5.

[0010] Furthermore, the CBR module of the amplification and fusion module extracts feature t4 to obtain feature s5, and inputs feature s5 into the first PSFF module for feature fusion. The Pixshuffle module upsamples feature s2 output by the third PSFF module, aligns the number of channels to obtain feature s1, and inputs feature s1 into the ORM module for refinement.

[0011] Each PSFF module is the output of the next PSFF module and two features and three inputs. First, the output of the next PSFF module is s i+2 As the main feature, t i+1 and t i Unified into s i+2With the same size and number of channels, they are added and sent to the SE channel attention module to further extract effective features, and finally combined with the main feature s i+2 Multiply to obtain the enhanced feature h1;

[0012] The second article uses deep features t i+1 As the main feature, s i+2 and t i Unified into t i+1 With the same size and number of channels, they are added and sent to the SE channel attention module to further extract effective features, and finally combined with the main feature t i+1 Multiply to obtain the enhanced feature h2;

[0013] The third article is based on shallow characteristics t i As the main feature, s i+2 and t i+1 Unified into t i With the same size and number of channels, they are added and sent to the SE channel attention module to further extract effective features, and finally combined with the main feature t i Multiply to obtain the enhanced feature h3;

[0014] Then, the features h1 and h2 are Pixshuffled upsampled to the same size as h3, and then they are channel-fused with h3. The fused features are sent to a 1×1 convolutional layer for channel reduction to obtain the final fused feature s i+1 , its specific expression is as follows:

[0015]

[0016] Among them, D(*) represents downsampling, SE(*) represents SE attention mechanism, P(*) represents Pixshuffle upsampling, C(*) represents channel fusion, and Conv1 represents 1×1 convolution.

[0017] Furthermore, when the feature refinement module focuses on the edge, p i+1 First, it is upsampled by bicubic interpolation, then it is dilated, and finally it is subtracted from the upsampled result to get the p when focusing on the edge. i+1 Input; when the module focuses on the target, p i+1 Directly obtain the p of the attention target through bicubic interpolation upsampling i+1 Input, p i+1 The input can be expressed as:

[0018]

[0019] DL(*) represents the dilation algorithm, which is a basic morphological operation; UP represents bicubic interpolation upsampling.

[0020] The p i+1 With low-level S i Multiply them, then send them to a 3×3 convolution to further extract features, and then multiply them with the learnable parameter α, and add the result to f i+1 Or the results after S5 alignment of size and number of channels are fused; then sent to CBR to further extract features and output enhanced features f i Finally, f i Send it to the 3×3 convolution to change the number of channels to get the predicted image p i ; The whole process can be expressed as:

[0021]

[0022] Among them, CBR(*) is Conv+BN+ReLU, C(*) represents channel fusion, UP represents bicubic interpolation upsampling, α represents a learnable parameter, and Conv3 represents 3×3 convolution.

[0023] Preferably, p i (i=1.2.3.4.5) Adjust the size to the same size as the true value image and calculate the loss. Use the binary cross entropy loss function and the intersection-over-union loss function to guide network training. The expressions are as follows:

[0024]

[0025] Among them, L iou represents the Intersection over Union (IoU) loss, L bce represents the binary cross entropy (BCE) loss, L represents the number of pixels, G represents the ground truth map of the disguised object, and δ represents the indicator function;

[0026] Combining the above two loss functions, we get the total loss function, which is expressed as follows:

[0027]

[0028] in,

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] The positioning module SE-ASPP (Compressed and Excited Atrous Spatial Convolution Pooling Pyramid) in this application can help the model focus on more important features and suppress irrelevant or redundant information by re-weighting the feature channels, thereby improving the feature selection capability and effectively solving the problem of inaccurate positioning of small targets; the Pixshuffle-based upsampling method is applied in the camouflaged target detection task, which loses less information than the bicubic interpolation upsampling method and restores better details.

[0031] The MFM module proposed in this application adopts a multi-feature fusion method. A PSFF module not only fuses adjacent features, but also fuses the features output by the subsequent PSFF module, establishes a connection between two adjacent PSFF modules, and eliminates the huge difference between deep semantic features and shallow detail features through the fusion of three features, thereby improving the accuracy of the subsequent refinement stage; in the feature refinement stage, the expansion algorithm is used to focus on the target and the edge at the same time, so as to accurately separate the highly similar foreground and background. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is a schematic diagram of the PAFNet network structure of the present invention;

[0033] Figure 2 It is a schematic diagram of the SE-ASPP structure of the present invention;

[0034] Figure 3 It is a schematic diagram of the PSFF structure of the present invention;

[0035] Figure 4 It is a schematic diagram of the ERM structure of the present invention;

[0036] Figure 5 It is a schematic diagram of the ORM structure of the present invention;

[0037] Figure 6 The figure is a schematic diagram comparing the results of an embodiment of the present invention with those of other methods. DETAILED DESCRIPTION

[0038] The specific implementation of the present invention is further described in detail below with reference to the accompanying drawings and specific examples.

[0039] like Figure 1As shown in the figure, a camouflaged target detection method based on Pixshuffle feature amplification and multi-feature fusion is used for camouflaged target detection PAFNet network detection, wherein the network includes a search and positioning module (SPM), a magnification and fusion module (MFM) and a feature refinement module (FRM), wherein the search and positioning module is composed of a backbone network PVTv2 and an improved version of a dilated spatial convolutional pooling pyramid (SE-ASPP), the magnification and fusion module is composed of a CBR module, three Pixshuffle Feature Fusion (PSFF) modules and a Pixshuffle module, and the feature refinement module is composed of three Edge Refinement Module (ERM) modules, an Object Refinement Module (ORM) module and a CBR module.

[0040] The backbone network in the search and positioning module (SPM) uses PVTv2-B3 to extract features of different scales. 3×H×W , and extract features of different scales through four Block modules in turn Then the deepest feature t4 is input into the SE-ASPP module for rough positioning to obtain the initial prediction map p5. In the camouflaged target detection task, the initially generated positioning map is crucial. If the initial positioning is wrong, then the subsequent segmentation process must be wrong. Therefore, this application improves the ASPP module, and the improved module is called SE-ASPP (Squeeze and Excitation-Atrous SpatialPyramidPooling).

[0041] The five branches of the improved SE-ASPP are all added with SE channel attention modules, which selectively emphasize useful features and suppress useless features through the SE module, thereby extracting more effective features to obtain more accurate target positioning prediction. Figure 2 As shown, the specific method of obtaining the initial prediction map p5 by rough positioning is as follows:

[0042] The deepest level feature t4 is input into five different branches in the SE-ASPP module, which includes three 3×3 hole convolution layers with expansion rates of 6, 12 and 18, a pooling layer and a 1×1 convolution layer. The hole convolution has different receptive fields and can extract features at multiple scales and capture targets of different sizes. The features obtained on the five different branches are respectively input into the SE channel attention module for feature extraction. Finally, the enhanced features of the five branches are fused, and then the channel is reduced through the 1×1 convolution layer to obtain the initial prediction map p5 for coarse positioning.

[0043] Furthermore, the CBR module of the amplification and fusion module (MFM) extracts feature t4 to obtain feature s5, and inputs feature s5 into the first PSFF module for feature fusion. The Pixshuffle module upsamples feature s2 output by the third PSFF module, aligns the number of channels to obtain feature s1, and inputs feature s1 into the ORM module for refinement.

[0044] The PSFF module (Pixshuffle Feature Fusion) aims to improve the resolution of feature maps and fuse them to make the feature maps clearer, which is beneficial for target segmentation in the subsequent refinement stage. Figure 3 As shown, each PSFF module is the output of the next PSFF module and two features and three inputs. First, the output of the next PSFF module is s i+2 As the main feature, t i+1 and t i Unified into s i+2 With the same size and number of channels, they are added and sent to the SE channel attention module to further extract effective features, and finally combined with the main feature s i+2 Multiply to obtain the enhanced feature h1;

[0045] The other two branches are the same as the first branch, but the main features are different. The second branch is based on the deep feature t i+1 As the main feature, s i+2 and t i Unified into t i+1 With the same size and number of channels, they are added and sent to the SE channel attention module to further extract effective features, and finally combined with the main feature t i+1 Multiply to obtain the enhanced feature h2;

[0046] The third article is based on shallow characteristics t i As the main feature, s i+2 and t i+1 Unified into t i With the same size and number of channels, they are added and sent to the SE channel attention module to further extract effective features, and finally combined with the main feature t i Multiply to obtain the enhanced feature h3;

[0047] Then, the features h1 and h2 are Pixshuffled upsampled to the same size as h3, and then they are channel-fused with h3. The fused features are sent to a 1×1 convolutional layer for channel reduction to obtain the final fused feature s i+1 , its specific expression is as follows:

[0048]

[0049] Among them, D(*) represents downsampling, SE(*) represents SE attention mechanism, P(*) represents Pixshuffle upsampling, C(*) represents channel fusion, and Conv1 represents 1×1 convolution.

[0050] Furthermore, the feature refinement module (FRM) aims to further refine the rough prediction map by alternately paying attention to the target and the edge to obtain the final prediction map, which includes three ERM modules (Edge refinement modules) and one ORM (Object refinement module). The ERM module focuses on the edge, and the ORM module focuses on the target itself.

[0051] like Figure 4 and 5 As shown, when the feature refinement module focuses on the edge, p i+1 First, it is upsampled by bicubic interpolation, then it is dilated, and finally it is subtracted from the upsampled result to get the p when focusing on the edge. i+1 Input; when the module focuses on the target, p i+1 Directly obtain the p of the attention target through bicubic interpolation upsampling i+1 Input, p i+1 The input can be expressed as:

[0052]

[0053] DL(*) represents the dilation algorithm, which is a basic morphological operation; UP represents bicubic interpolation upsampling.

[0054] The p i+1 With low-level S i Multiply them, then send them to a 3×3 convolution to further extract features, and then multiply them with the learnable parameter α, and add the result to f i+1 Or the results after S5 alignment of size and number of channels are fused; then sent to CBR to further extract features and output enhanced features f i Finally, f i Send it to the 3×3 convolution to change the number of channels to get the predicted image p i ; The whole process can be expressed as:

[0055]

[0056] Among them, CBR(*) is Conv+BN+ReLU, C(*) represents channel fusion, UP represents bicubic interpolation upsampling, α represents a learnable parameter initialized to 1, and Conv3 represents 3×3 convolution.

[0057] The p i (i=1.2.3.4.5) Adjust the size to the same size as the true value image and calculate the loss. p1 is the final output image. The binary cross entropy loss function and intersection-over-union loss function are used to guide network training. The expressions are as follows:

[0058]

[0059] Among them, L iou represents the Intersection over Union (IoU) loss, L bce represents the binary cross entropy (BCE) loss, L represents the number of pixels, G represents the ground truth map of the disguised object, and δ represents the indicator function;

[0060] Combining the above two loss functions, we get the total loss function, which is expressed as follows:

[0061]

[0062] in,

[0063] Example:

[0064] 1) Dataset

[0065] To experimentally analyze the proposed method, PAFNet is tested on four benchmarks: CHAMELEON, CAMO, COD10K, and NC4K. CAMO contains a total of 1250 images, each with at least one camouflaged object, 1000 images as training set, and 250 images as test set. CAMO involves various challenging scenarios such as object appearance, background clutter, shape complexity, small objects, object occlusion, multiple objects, and distraction. CHAMELEON searches 76 images with camouflaged animals as keywords from Google search. COD10K is the largest benchmark dataset, containing 5066 camouflaged object images, including 3040 images as training set and 2026 images for testing, and it includes 5 superclasses and 69 subclasses. NC4K is the test dataset used in camouflaged object segmentation, with a total of 4121 images.

[0066] 2) Evaluation Metrics

[0067] The commonly used evaluation indicators in camouflaged target detection are as follows: structural measure (S α ), Adaptive E-measure F-measure, weighted F-measure Mean absolute error (M). Structural measurement S α Aiming to evaluate the structural information of the prediction, E-measure focuses more on pixel-level matching and image-level statistics, while F-measure focuses more on the comprehensive measurement of the prediction image precision and recall rate. It can provide more than traditional F β For more reliable evaluation results, the mean absolute error (M) is calculated as element-wise difference.

[0068] 3) Experimental details

[0069] The PyTorch framework is used to implement the model of this application, and Intel(R) Xeon(R) Silver 4114CPU@2.20GHz (total memory capacity 64G) is used for training and testing. For the training process, 4040 camouflaged target images from CAMO and COD10K are used for training, and the input image size is resized to a resolution of 384×384. The optimizer is a stochastic gradient descent (SGD) optimizer with a momentum of 0.9 and a weight decay of 5×10 -4 In addition, the batch size of the trained model is 16, the training epoch is 150, and the base learning rate is 0.001. For testing, the input images are resized to 384×384 for model inference and then resized to the original size for evaluation.

[0070] 4) Comparison with other methods

[0071] Table1 QUANTITATIVE COMPARISON BETWEEN OUR PROPOSED METHOD AND OTHERSTATE-OF-THE-ART METHODS ON FOUR BENCHMARKS.THE TOP RESULTS ARE HIGHLIGHTEDIN RED

[0072]

[0073] To demonstrate the effectiveness of the model, five evaluation metrics are used to compare the model with existing methods on four datasets. The existing methods include SINet, PFNet, SINet-V2, BGNet, UGTR, ZoomNet, CFANet, FEDER, EAMNet, FSPNet, and SARNet.

[0074] Quantitative analysis: From Table 1, we can see that the model achieved the best performance on all four datasets and was much better than other methods. In particular, all indicators on the CHAMELEON dataset exceeded the second best result by an average of 1.5%, and all indicators on the CAMO dataset exceeded the second best result by an average of 1.3%.

[0075] Qualitative analysis: Visual results compared with other methods are shown in Figure 6 As shown in the figure. From the results, we can see that the model can identify objects under various camouflages, such as background matching, small targets, and subtle target edges. For subtle parts of the target, the model can locate them well and accurately segment them. Taking the last row of insect antennae as an example, other methods will have unclear edge segmentation during segmentation, resulting in incomplete segmentation of the antennae, but our method can accurately segment the edges of the target, so our segmentation map is more accurate and complete than other methods. For background color and target color matching, our model can locate the target well and accurately segment them.

[0076] 5) Ablation experiment

[0077] A series of ablation experiments are conducted to demonstrate the effectiveness of the three components, namely PSFF, SE-ASPP and ERM (ORM), and ablation experiments are performed on the Pixshuffle upsampling, SE-ASPP improvement, and MFM connection method in this paper.

[0078] Table 2ABLATION STUDY ON PPRNET MODULE,"B"DENOTES BASIC MODEL.THEBEST RESULTS ARE HIGHLIGHTED IN RED.

[0079]

[0080] Effectiveness of PSFF: The PSFF module is used to improve feature resolution and perform feature fusion, which helps the model combine feature information at different levels and subsequent accurate segmentation. In Table 2, by comparing the experimental results of A and B on four data sets, all the effects of B are better than A, proving that PSFF can improve the performance of the model.

[0081] Effectiveness of SE-ASPP: The SE-ASPP module adds the SE attention mechanism to ASPP, which can further extract effective features and make the initial prediction map positioning more accurate. In Table 2, by comparing the experimental results of A and C on four data sets, all the results of C are better than A, proving that SE-ASPP can improve the performance of the model.

[0082] Effectiveness of ORM (ERM): The ORM (ERM) module further refines the rough prediction map by paying attention to the target and the edge alternately to obtain the final prediction map. In Table 2, by comparing the experimental results of A and D on four data sets, all the results of D are better than A, proving that ORM (ERM) can improve the performance of the model.

[0083] Table 3ABLATION STUDY ON PIX-SHUFFLE UPSAMPLING,"B"DENOTES BASICMODEL.Bicubic DENOTES AN INTERPOLATION UPSAMPLING.THE BEST RESULTS AREHIGHLIGHTED IN RED.

[0084]

[0085] Effectiveness of Pix-shuffle upsampling: Pix-shuffle is used to upsample features, which is conducive to the recovery of detail information. In Table 3, by comparing the experimental results of B and C on four data sets, all the effects of C are better than B, proving the accuracy of introducing Pixshuffle upsampling compared to ordinary upsampling.

[0086] Table 4ABLATION STUDY ON Connection method of MFM,"B"DENOTES BASICMODEL THE BEST RESULTS ARE HIGHLIGHTED IN RED.

[0087]

[0088] Effectiveness of MFM connection method: The MFM module adopts a connection method that forwards the information of each module. This connection method can effectively eliminate the huge difference between deep semantic features and shallow detail features, and improve the accuracy of the subsequent refinement stage. In Table 4, by comparing the experimental results of G and H on the CAMO and CHAMELEON datasets, all results of H are better than G, proving that the MFM connection method can improve the performance of the model.

[0089] Table 5 ABLATION STUDY ON SE-ASPP,"B"DENOTES BASIC MODEL.THE BESTRESULTS ARE HIGHLIGHTED IN RED.

[0090]

[0091] Effectiveness of SE-ASPP Improvement: The SE-ASPP module adds the SE attention mechanism to ASPP, which can further effectively extract features and make the initial prediction map positioning more accurate. In Table 5, by comparing the experimental results of E and F on the CAMO and CHAMELEON datasets, all results of F are better than E, proving that the improvement of SE-ASPP can improve the performance of the model.

[0092] The above description is only an implementation mode of the present invention. It is stated again that for ordinary technicians in this technical field, several improvements can be made to the present invention without departing from the principle of the present invention. These improvements are also included in the protection scope of the claims of the present invention.

Claims

1. A camouflaged target detection method based on Pixshuffle feature amplification and multi-feature fusion is used for camouflaged target detection PAFNet network detection, which is characterized by: The network includes a search and positioning module, an amplification and fusion module and a feature refinement module, wherein the search and positioning module is composed of a backbone network PVTv2 and an improved version of a dilated spatial convolutional pooling pyramid, the amplification and fusion module is composed of a CBR module, three Pixshuffle Feature Fusion modules and a Pixshuffle module, and the feature refinement module is composed of three Edge Refinement Module modules, an Object RefinementModule module and a CBR module.

2. The camouflaged target detection method based on Pixshuffle feature amplification and multi-feature fusion according to claim 1 is characterized in that: The backbone network in the search and positioning module uses PVTv2-B3 to extract features of different scales through the input image I∈R 3×H×W , and extract features of different scales through four Block modules in turn Then the deepest level feature t4 is input into the SE-ASPP module for coarse positioning to obtain the initial prediction map p5.

3. The camouflaged target detection method based on Pixshuffle feature amplification and multi-feature fusion according to claim 2 is characterized in that: The specific method of obtaining the initial prediction map p5 by rough positioning is as follows: The deepest level feature t4 is input into five different branches in the SE-ASPP module, which includes three 3×3 hole convolution layers with expansion rates of 6, 12 and 18, a pooling layer and a 1×1 convolution layer. The features obtained from the five different branches are respectively input into the SE channel attention module for feature extraction. Finally, the enhanced features of the five branches are fused, and then the channel is reduced through a 1×1 convolution layer to obtain the coarse positioning map p5.

4. The camouflaged target detection method based on Pixshuffle feature amplification and multi-feature fusion according to claim 1 is characterized in that: The CBR module of the amplification and fusion module extracts feature t4 to obtain feature s5, and inputs feature s5 into the first PSFF module for feature fusion. The Pixshuffle module upsamples feature s2 output by the third PSFF module, aligns the number of channels to obtain feature s1, and inputs feature s1 into the ORM module for refinement. Each PSFF module is the output of the next PSFF module and two features and three inputs. First, the output of the next PSFF module is s i+2 As the main feature, t i+1 and t i Unified into s i+2 With the same size and number of channels, they are added together and sent to the SE attention mechanism to further extract effective features, and finally combined with the main feature s i+2 Multiply to obtain the enhanced feature h1; The second article uses deep features t i+1 As the main feature, s i+2 and t i Unified into t i+1 With the same size and number of channels, they are added and sent to the SE channel attention module to further extract effective features, and finally combined with the main feature t i+1 Multiply to obtain the enhanced feature h2; The third article is based on shallow characteristics t i As the main feature, s i+2 and t i+1 Unified into t i With the same size and number of channels, they are added and sent to the SE channel attention module to further extract effective features, and finally combined with the main feature t i Multiply to obtain the enhanced feature h3; Then, the features h1 and h2 are Pixshuffled upsampled to the same size as h3, and then they are channel-fused with h3. The fused features are sent to a 1×1 convolutional layer for channel reduction to obtain the final fused feature s i+1 , its specific expression is as follows: Among them, D(*) represents downsampling, SE(*) represents SE attention mechanism, P(*) represents Pixshuffle upsampling, C(*) represents channel fusion, and Conv1 represents 1×1 convolution.

5. The camouflaged target detection method based on Pixshuffle feature amplification and multi-feature fusion according to claim 1 is characterized in that: When the feature refinement module focuses on the edge, p i+1 First, it is upsampled by bicubic interpolation, then it is dilated, and finally it is subtracted from the upsampled result to get the p when focusing on the edge. i+1 Input; when the module focuses on the target, p i+1 Directly obtain the p of the attention target through bicubic interpolation upsampling i+1 Input, p i+1 The input can be expressed as: Where DL(*) represents the dilation algorithm, and UP represents bicubic interpolation upsampling; The p i+1 With low-level S i Multiply them, then send them to a 3×3 convolution to further extract features, and then multiply them with the learnable parameter α, and add the result to f i+1 Or the results after S5 alignment of size and number of channels are fused; then sent to CBR to further extract features and output enhanced features f i Finally, f i Send it to the 3×3 convolution to change the number of channels to get the predicted image p i ; The whole process can be expressed as: Among them, CBR(*) is Conv+BN+ReLU, C(*) represents channel fusion, UP represents bicubic interpolation upsampling, α represents a learnable parameter, and Conv3 represents 3×3 convolution.

6. The camouflaged target detection method based on Pixshuffle feature amplification and multi-feature fusion according to claim 1 is characterized in that: The p i (i=1.2.3.4.5) Adjust the size to the same size as the true value image and calculate the loss. Use the binary cross entropy loss function and the intersection-over-union loss function to guide network training. The expressions are as follows: Among them, L iou represents the Intersection over Union (IoU) loss, L bce represents the binary cross entropy (BCE) loss, L represents the number of pixels, G represents the ground truth map of the disguised object, and δ represents the indicator function; Combining the above two loss functions, we get the total loss function, which is expressed as follows: in,

Images