A camouflaged target detection method based on attention mechanism and convolutional neural network

Abstract

The present invention discloses a camouflaged target detection method based on an attention mechanism and a convolutional neural network, belonging to the field of camouflaged target detection. Specifically, the method comprises: inputting a training set image into a backbone network to extract multi-scale features containing the camouflaged target image; inputting the features output by Stage 3, Stage 4, and Stage 5 of the backbone network into a position-aware recurrent convolution module to output global features; utilizing an edge extraction module to extract edge contour information of the camouflaged target to obtain an edge prediction map #imgabs0#, and performing boundary supervision on #imgabs2# using the edge label #imgabs1# of the camouflaged target; fusing the obtained global features with the edge contour information, and then performing multi-scale feature fusion to obtain a multi-scale aggregated feature #imgabs3#, obtaining a camouflaged target prediction map, and performing deep supervision on the camouflaged target using the binary label map of the camouflaged target. The camouflaged target detection method of the present invention can fully perceive the camouflaged target and refine its boundary contour, thereby improving the detection performance of the camouflaged target.

Description

Technical Field

[0001] The present invention relates to a camouflaged target detection method based on an attention mechanism and a convolutional neural network, and belongs to the field of camouflaged target detection. Background Art

[0002] Many creatures in nature possess the ability to camouflage. Chameleons can adjust their color to suit their surroundings; lions hide in the grass while waiting for prey to approach; and butterflies lie motionless on tree trunks of similar color to their own to avoid harm from predators. Biologists call this type of camouflage background matching: to avoid recognition, animals attempt to change their color to perfectly blend in with their surroundings. Therefore, compared to general object detection and salient object detection, where the object and background are clearly distinct and generally easy for the human eye to distinguish, the high similarity between the camouflaged object and the background in camouflaged object detection makes it more challenging.

[0003] The boundary between the boundary of the camouflaged target and the background is quite fuzzy and difficult to distinguish. Without introducing additional prior information, it is difficult to accurately locate the camouflaged object. The paper "Camouflaged object segmentation with distraction mining. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, June 2021" introduces the concept of PFNet interference removal into the camouflaged object segmentation task, and develops a new mining strategy for the discovery and removal of interference areas to assist in the segmentation of camouflaged objects. However, it does not focus on the boundary information of the camouflaged object and cannot accurately segment the complete boundary of the camouflaged object; the paper "Camouflaged object detection. In CVPR, 2020." SINet proposed that the receptive field module (RFB) be used to increase the receptive field and promote the segmentation of disguised targets. However, RFB can only enhance the local receptive field and cannot obtain global features, while enhancing the global receptive field cannot obtain global context information. The Chinese invention patent with patent number CN113468996A discloses a disguised object detection method based on edge refinement. Although it considers edge prior information, it only considers global average pooling in the attention mechanism of the edge refinement module, resulting in the loss of a large amount of available different frequency information. In summary, the capabilities of existing disguised target detection algorithms need to be improved. Summary of the Invention

[0004] In response to the above-mentioned problems, the present invention provides a camouflaged target detection method based on an attention mechanism and a convolutional neural network. By using an edge extraction module to effectively extract edge contour information, the frequency channel attention component FcaNet introduced in the edge enhancement module extracts different frequency component information through two-dimensional discrete cosine transform DCT and combines them, which can capture diverse feature information and fuse it with the extracted global features to enhance boundary representation. A multi-scale attention mechanism is introduced into the camouflaged target detection network model to effectively aggregate multi-scale features, thereby achieving comprehensive perception of camouflaged targets and refining their boundary contours, thereby improving the detection performance of camouflaged targets.

[0005] The technical solutions adopted by the present invention to solve the above technical problems are as follows:

[0006] A method for detecting camouflaged targets based on an attention mechanism and a convolutional neural network includes the following steps:

[0007] S1. Divide the image dataset of the camouflaged target into a training set and a test set;

[0008] S2, the training set images Input into the backbone network of the pre-built camouflaged target detection network model to extract multi-scale features containing camouflaged target images , The resolutions are ;

[0009] S3, the features output by the feature extraction layers Stage3, Stage4, and Stage5 of the backbone network They are respectively input into the position-aware recurrent convolution module PARCM to output global features;

[0010] S4. Use the edge extraction module to extract the edge contour information of the camouflaged target , and then get the edge prediction map , and by disguising the edge labels of the target right Conduct border surveillance;

[0011] S5: After effectively fusing the global features obtained in step S3 with the edge contour information obtained in step S4, multi-scale feature fusion is performed to obtain multi-scale aggregated features. ;

[0012] S6, the multi-scale aggregation features obtained in step S5 The camouflaged target prediction map is obtained through processing, and deep supervision is performed on it through the binary label map of the camouflaged target;

[0013] S7, test set images As the input of the trained disguised target detection network model, the final disguised target detection result is obtained.

[0014] Furthermore, the backbone network uses the EfficientNet-B4 model in the EfficientNet series to extract multi-scale features containing the disguised target image.

[0015] Furthermore, the position-aware circular convolution module PARCM in step S3 includes a position-aware circular convolution component ParC and a channel attention component, wherein the position-aware circular convolution component ParC uses global circular convolution GCC to extract global features.

[0016] Furthermore, in step S4, the edge extraction module is used to extract the edge contour information of the camouflaged target, and then the edge prediction map is obtained. , and by disguising the edge labels of the target right The specific contents of boundary supervision include: using the edge extraction module EEM to extract the low-level features output by the feature extraction layer Stage2 of the backbone network High-level semantic features output by the feature extraction layer Stage5 The edge contour information of the camouflaged target is extracted by fusion, and the output of the edge extraction module EEM is normalized by the Sigmoid function to obtain a binary image ,Will Upsample four times to get the edge prediction map , by disguising the edge labels of the target right For boundary supervision, the edge loss function used is .

[0017] Furthermore, a position embedding strategy is introduced in the position-aware recurrent convolution module PARCM.

[0018] Furthermore, nonlinear characteristics are introduced into the channel attention component of the position-aware recurrent convolution module PARCM through a feedforward neural network (FFN), and a channel attention mechanism SE Block is added after FFN to highlight key channels.

[0019] Furthermore, residual connections are introduced in the position-aware recurrent convolution module PARCM.

[0020] Furthermore, after the global features obtained in step S3 are effectively fused with the edge contour information obtained in step S4, multi-scale feature fusion is then performed to obtain multi-scale aggregated features. The specific contents include: using the edge enhancement module ERM to transform the edge contour information from the edge extraction module EEM Global features output by the position-aware recurrent convolution module PARCM Fusion to obtain features Then, the multi-scale attention mechanism MSAM is introduced into the multi-scale fusion module MSFM to perform multi-scale feature fusion, where the feature The multi-scale feature fusion module MSFM is used to fuse the multi-scale features to obtain the features. , then the feature and features Multi-scale aggregation features are obtained by feature fusion through the multi-scale fusion module MSFM For disguised datasets, the sizes of disguised objects usually vary. The multi-scale attention mechanism MSAM has strong adaptability to disguised targets of different scales and can achieve effective fusion of multi-scale features.

[0021] Furthermore, the frequency channel attention component FcaNet is introduced into the edge enhancement module ERM. The two-dimensional discrete cosine transform (DCT) is used to extract information of different frequency components and combine them. Compared with the single global average pooling operation, it can capture diverse feature information and enhance edge details.

[0022] Furthermore, the structured loss function selected for deep supervision of the binary label map of the disguised target in step S6 is the weighted binary cross loss and weighted intersection-over-union loss ;

[0023] Furthermore, the total loss function of the disguised target detection model is The formula is as follows:

[0024] ;

[0025] ;

[0026] ;

[0027] in, represents the structured loss, represents the edge loss, and are the weight factors of structured loss and edge loss respectively; is the predicted camouflaged target saliency map, Represents the predicted edge contour of the camouflaged target; represents the saliency label of the disguised target, edge labels representing disguised targets; is the weighted binary cross loss in the structured loss, is the weighted intersection-over-union loss in the structured loss, is the edge loss obtained using the Dice coefficient.

[0028] Compared with the prior art, the technical solution of the present invention has the following technical effects:

[0029] The present invention introduces additional edge prior information and performs deep supervision, introduces the frequency channel attention component FcaNet into the edge enhancement module, extracts different frequency component information through two-dimensional discrete cosine transform DCT and combines them. Compared with a single global average pooling operation, it can capture diverse feature information and fuse it with the global features extracted by global circular convolution GCC to enhance boundary representation.

[0030] The position-aware recurrent convolution module of the camouflaged target detection network model in the present invention can effectively extract global features and obtain a global receptive field by introducing global recurrent convolution GCC, thereby compensating for the problem of excessive locality and insufficient globality of the convolutional neural network, thereby obtaining global context information; at the same time, by introducing a position embedding strategy in the position-aware recurrent convolution module, position information is injected into the output feature map to ensure the sensitivity of the output features to spatial positions, and a channel attention mechanism is introduced to highlight key channels; and the present invention introduces a multi-scale attention mechanism in the camouflaged target detection network model, which can effectively fuse multi-scale context information, thereby improving the camouflaged target detection performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 This is an overall flow chart of the implementation of the camouflaged target detection method based on the attention mechanism and convolutional neural network of the present invention;

[0032] Figure 2 This is a flowchart for implementing the position-aware circular convolution module PARCM according to an embodiment of the present invention;

[0033] Figure 3 Schematic diagram of the calculation process of the global circular convolution GCC in the position-aware circular convolution module PARCM according to an embodiment of the present invention;

[0034] Figure 4 This is a flow chart for implementing the edge extraction module EEM according to an embodiment of the present invention;

[0035] Figure 5 This is a flowchart of an implementation of the edge enhancement module ERM according to an embodiment of the present invention;

[0036] Figure 6This is a flowchart for implementing the frequency channel attention component FcaNet according to an embodiment of the present invention;

[0037] Figure 7 This is a flowchart for implementing the multi-scale attention mechanism MSAM described in an embodiment of the present invention;

[0038] Figure 8 This is a flowchart for implementing the multi-scale fusion module MSFM according to an embodiment of the present invention;

[0039] Figure 9 The image segmentation results of five mainstream camouflaged target segmentation models and the method of the present invention on different camouflaged target data sets in the comparative test of the embodiment of the present invention are compared. DETAILED DESCRIPTION

[0040] To make the purpose, technical solutions and advantages of the embodiments of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the specific embodiments of the present application and the corresponding drawings. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of them.

[0041] Figure 1 This is the overall flow chart of the implementation of the camouflaged target detection method based on the attention mechanism and convolutional neural network of the present invention, as shown in the figure: Figure 1 As shown, the camouflaged target detection network model of the present invention includes a backbone network, a position-aware recurrent convolution module PARCM, an edge extraction module EEM, an edge enhancement module ERM, and a multi-scale fusion module MSFM. The backbone network uses the EfficientNet-B4 model from the EfficientNet series to extract multi-scale features containing camouflaged target images. Since the input images of the camouflaged target training set vary in size, EfficientNet is selected as the backbone network, which can balance the three dimensions of network depth, network width and resolution, thereby effectively improving the feature extraction capability; the position-aware circular convolution module PARCM is used to extract global features and enhance the receptive field; the edge extraction module EEM is used for edge extraction and outputs edge contour information; the edge enhancement module ERM is used to effectively fuse the edge contour information output by the edge extraction module EEM with the features enhanced by the receptive field of the position-aware circular convolution module PARCM to enhance the boundary representation; the multi-scale fusion module MSFM effectively fuses multi-scale features by introducing the multi-scale attention mechanism MSAM.

[0042] The method for detecting disguised targets based on the attention mechanism and convolutional neural network described in the present invention includes the following contents:

[0043] Step 1: Divide the image dataset of the camouflaged target into a training set and a test set

[0044] Step 2: The training set images Input into the pre-built camouflaged target detection network model backbone network EfficientNet to extract multi-scale features containing camouflaged target images , The resolutions are ,in and Represent the height and width of the feature map respectively, Indicates the resolution scaling factor.

[0045] Step 3: The features output by the feature extraction layers Stage3, Stage4, and Stage5 of the backbone network Input into the position-aware recurrent convolution module PARCM to output global features

[0046] In this embodiment, the features output by the feature extraction layers Stage3, Stage4, and Stage5 of the backbone network are Input into the position-aware recurrent convolution module PARCM output respectively The implementation flow chart of the position-aware circular convolution module PARCM according to the embodiment of the present invention is as follows: Figure 2 As shown, the position-aware recurrent convolution module (PARCM) consists of a position-aware recurrent convolution component (ParC) and a channel attention component. The ParC uses global recurrent convolutions (GCCs), which include horizontal GCC-H and vertical GCC-V. The combined use of GCC-H and GCC-V extracts global features from all input positions and achieves a global receptive field. The convolution kernel size in the GCCs is consistent with the input, and a position embedding strategy is used to inject position information into the output feature map, ensuring the sensitivity of the output features to spatial position and reducing the interference of recurrent convolutions on spatial structure. Furthermore, the channel attention component introduces nonlinearity through a feedforward neural network (FFN), and a channel attention mechanism (SE Block) is added after the FFN to highlight key channels and suppress feature channels that are less useful for the task. Furthermore, the position-aware recurrent convolution module (PARCM) incorporates residual connections, enabling it to simultaneously perceive global and local contextual information. Among them, PW-Conv is DW convolution, which is used to adjust the input dimension. In addition, Pre-Norm uses batch normalization operation.

[0047] Figure 3 Schematic diagram of the calculation process of the global circular convolution GCC in the position-aware circular convolution module PARCM according to an embodiment of the present invention, as shown in FIG. Figure 3 As shown, input Through the calculation process of GCC-H and GCC-V, for simplicity, it is assumed that the input There is only one channel, and the corresponding shape is , The pixel at position passes through the output of GCC-V It can be calculated by the following formula:

[0048]

[0049] in It is a basic position embedding information with dimension ; Through the bilinear interpolation function , get instance position embedding information , the dimension is ; Through the vertical expansion function , after the input vector along Direction Copy times, generate The PE matrix of size is The dimension is , so that the position embedding PE can flexibly adapt to input features of different sizes; embed the instance position information With input Superposition to obtain features ; The feature Along Direction stacking, the feature dimension obtained is , which can realize standard convolution with global receptive field and parameter sharing; next, by constructing the convolution kernel parameters , also by a size of Basic amount , through the bilinear interpolation function , and the size is of , at this time the spatial receptive field of the convolution kernel is , with the spatial receptive field of the GCC-H convolution kernel , approximate a global cover; Input with embedded location information Perform a standard convolution operation, i.e. In this process, the size of GCC-V is Output feature coordinates of the convolution kernel , which actually corresponds to the coordinate range on the input feature ,in is a 0 to The values ​​between Corresponding to the relative coordinates in the local neighborhood covered by the convolution kernel, the modulo operation is equivalent to a loop operation, which corresponds to the global circular convolution. The specific code implementation can be achieved through Similarly, GCC-H is The output at can be expressed as:

[0050]

[0051] in , The expansion function expands the input vector vertically. In summary, the receptive fields output by GCC-H and GCC-V can cover the same columns and rows of all input positions. The combined use of GCC-H and GCC-V can extract global features from all input positions and obtain a global receptive field.

[0052] Step 4: Use the edge extraction module EEM to extract the edge contour information of the camouflaged target , and then get the edge prediction map , and by disguising the edge labels of the target right Conduct border monitoring

[0053] Figure 4 This is a flow chart of the implementation of the edge extraction module EEM according to an embodiment of the present invention, as shown in FIG. Figure 4 As shown, the edge extraction module EEM obtains the edge prediction map The specific contents include: first use The convolution transforms the features and features The number of channels is compressed, and then the features are Upsample to the same feature After the two are of the same size, they are spliced ​​together and then The convolution is fused and finally a The convolution and Sigmiod function are normalized to obtain the edge prediction map , where the edge loss function is selected Due to the low-level features There is a lot of noise in the image, so the edge extraction module EEM introduces high-level semantic features. As an auxiliary, the main function of the edge extraction module EEM is to provide valuable edge priors for subsequent segmentation, so that the model can better segment the edge contours of the camouflaged target.

[0054] Step 5: The global features obtained in step 3 are effectively fused with the edge contour information obtained in step 4 to obtain edge enhancement features.

[0055] Figure 5 This is a flow chart of the implementation of the edge enhancement module ERM according to an embodiment of the present invention, as shown in FIG. Figure 5 As shown, the edge enhancement module ERM obtains the edge enhancement feature The specific contents include: first, the edge contour map Downsample to the output features of the position-aware circular convolution module PARCM Same size, then with Perform element-wise multiplication and combine the result with Add element by element, followed by a The convolution layer is then normalized and activated with ReLU. The frequency channel attention component FcaNet is introduced, and the input and output of the frequency channel attention component are added element by element through the residual connection to obtain the boundary reinforcement feature. .

[0056] The edge reinforcement module ERM introduces the frequency channel attention component FcaNet, Figure 6 This is the implementation flow chart of the frequency channel attention component FcaNet, as shown Figure 6 As shown in the figure, FcaNet extracts different frequency component information through two-dimensional discrete cosine transform (DCT) and combines them. Compared with a single global average pooling operation, it can capture diverse feature information and enhance edge details.

[0057] The two-dimensional discrete cosine transform DCT is calculated as follows:

[0058]

[0059] In the formula are the width and height of the image, for The value of the pixel at , the frequency component It can be regarded as the weighted sum of each input point in the image, and the cosine part is equivalent to the weight. The above formula can realize the conversion from spatial domain to frequency domain. for

[0060]

[0061] Since global average pooling can be expressed as , so the lowest frequency component and Directly proportional.

[0062] make , the two-dimensional discrete cosine transform inverse transform can be expressed as:

[0063]

[0064] From the above formula, we can get that the original image signal is composed of the global average pooling term GAP and other frequency components. The pixel value at All frequency components can be Previous channel attention components such as SENet only considered GAP (the lowest frequency component information), resulting in the loss of a large amount of available information. FcaNet extracts different frequency component information through two-dimensional discrete cosine transform (DCT) and effectively combines them to obtain diverse feature information and enhance feature representation and edge details.

[0065] Step 6: Enhance the edge features obtained in step 5 Perform multi-scale feature fusion to obtain multi-scale aggregation features

[0066] The multi-scale fusion module MSFM introduces the multi-scale attention mechanism MSAM. Figure 7 This is the implementation flow chart of the multi-scale attention mechanism MSAM, as shown Figure 7 As shown, MSAM consists of two branches. One branch uses a global average pooling layer and two point-by-point convolutions to obtain global context information, highlighting globally distributed large objects. The other branch maintains the original feature size and uses only two point-by-point convolutions to obtain local context information, preventing small objects from being overlooked. Finally, the two branches are concatenated and passed through a sigmoid function to obtain multi-scale channel attention coefficients. Features at different levels contribute differently to the task, and the fusion of multi-level features can complement each other to obtain a comprehensive feature representation. Figure 8 This is a flowchart for implementing the multi-scale fusion module MSFM according to an embodiment of the present invention, and its formula is as follows:

[0067]

[0068] in, It is a feature and The result after passing the multi-scale fusion module MSFM, Represents the multi-scale attention mechanism MSAM component, Represents low-level features, Represents high-level features, Represents high-level features that are upsampled twice , represents the element addition operation, Represents an element-wise multiplication operation.

[0069] The low-level features and high-level features are fused to obtain , then enter a The convolution layer then outputs the fused multi-scale features after batch normalization and ReLU activation function. In the specific implementation, this embodiment takes into account that low-level features with larger spatial resolution require more computing resources than high-level semantic features and contribute less to the performance of the model. Based on this observation, only high-level features are used. Perform feature fusion, where The feature fusion is obtained by multi-scale fusion module MSFM , then and The feature fusion is obtained by multi-scale fusion module MSFM ,Features of different scales contribute differently to the task, and the fusion of features at multiple scales through the ,multiscale fusion module MSFM can complement each other to obtain a comprehensive feature ,representation.

[0070] In step seven, the multi-scale aggregated features obtained in step six are convolved to obtain a disguised target prediction map, and deep supervision is performed on the disguised target through the binary label map of the disguised target.

[0071] The above obtained Through a Convolution obtains a single-channel grayscale image, and then an 8-fold upsampling is performed to obtain the final camouflaged target prediction image, and the binary label image of the camouflaged target is obtained. Perform deep supervision and calculate its structured loss. The selected loss function is binary cross loss. and weighted intersection-over-union loss The total loss function of the final model is:

[0072]

[0073]

[0074]

[0075] in, and is the weight factor of each loss. In the experimental simulation and Take 1 and 3 respectively, represents the structured loss, represents the edge loss, is the predicted camouflaged target saliency map, represents the predicted camouflaged target edge map, represents the saliency label of the disguised target, The edge labels representing the disguised targets.

[0076] Finally, the test set images As the input of the trained disguised target detection network model, the final disguised target detection result is obtained.

[0077] In order to verify the effectiveness of the present invention, the image segmentation results of five mainstream camouflaged target segmentation models UGTR, MGL-R, PFNet, EGNet, and SINet were tested on the camouflaged target datasets CAMO, CHAMELEON, COD10K, and NC4K, and compared with the camouflaged target prediction results of the present invention.

[0078] The experimental simulation content of the camouflaged target detection method based on the attention mechanism and convolutional neural network of the present invention is as follows:

[0079] The experimental platform uses a 64-bit Ubuntu system with a system version of 20.04.4, a GPU model of GeForceRTX 2080 Ti, Python 3.8 as the programming language, and PyCharm as the software development platform. The model is implemented using the deep learning framework Pytorch 1.4, and the input image is a 3-channel RGB image with a size of During the training phase, the batch size is 24 and the number of epochs is 40. The parameters are optimized using the AdaX optimizer. The initial learning rate is set to 3e-4, and the poly strategy is used to continuously adjust it with a power of 0.9. With the acceleration of three GeForce RTX 2080 Ti GPUs, the entire network training time is about 1 hour. Figure 9 The image segmentation results of five mainstream camouflaged target segmentation models and the method of the present invention on different camouflaged target data sets in the comparative test of the embodiment of the present invention are compared. Denotes the saliency label of the disguised target, and Our denotes the disguised target prediction result of the present invention. By comparison, it can be seen that the present invention can effectively detect the disguised target, refine the edge contour of the disguised target, and improve the accuracy of disguised target detection.

[0080] The above description is merely a preferred embodiment of the present invention and does not limit the present invention in any way. Any person skilled in the art who, without departing from the scope of the present invention, makes any equivalent substitution, modification, or other variation to the technical solution and technical contents disclosed in the present invention shall be deemed to be within the scope of the present invention and still fall within the scope of protection of the present invention.

Claims

1. A camouflaged target detection method based on attention mechanism and convolutional neural network, characterized in that: The specific steps include: S1. Divide the image dataset of the camouflaged target into a training set and a test set; S2. Input the training set images into the backbone network of the pre-built camouflaged target detection network model to extract the multi-scale features f containing the camouflaged target image. i ,i∈{1,2,3,4,5}; S3, the feature f output by the feature extraction layers Stage3, Stage4, and Stage5 of the backbone network i ,i∈{3,4,5} are respectively input into the position-aware recurrent convolution module PARCM to output global features; The position-aware circular convolution module PARCM in step S3 includes a position-aware circular convolution component ParC and a channel attention component, wherein the position-aware circular convolution component ParC uses a global circular convolution GCC to extract global features, and introduces a residual connection in the position-aware circular convolution module PARCM; S4, using the edge extraction module EEM to extract the edge contour information of the camouflaged target e , and then get the edge prediction map O e , and by disguising the target edge label GE to O e Conduct border surveillance; S5: After effectively fusing the global features obtained in step S3 with the edge contour information obtained in step S4, multi-scale feature fusion is performed to obtain the multi-scale aggregated feature f 345 ; S6, the multi-scale aggregation feature f obtained in step S5 345 The camouflaged target prediction map is obtained through processing, and deep supervision is performed on it through the binary label map of the camouflaged target; S7. Use the test set images as input to the trained camouflaged target detection network model to obtain the final camouflaged target detection results.

2. The method for detecting disguised targets based on an attention mechanism and a convolutional neural network according to claim 1, wherein: The backbone network uses the EfficientNet-B4 model in the EfficientNet series to extract multi-scale features of the disguised target image.

3. The method for detecting camouflaged targets based on an attention mechanism and a convolutional neural network according to claim 2, wherein: In step S4, the edge extraction module EEM is used to extract the edge contour information of the camouflaged target, and then the edge prediction image O is obtained. e , and by disguising the target edge label GE to O e The specific contents of boundary supervision include: using the edge extraction module EEM to fuse the low-level features f2 output by the feature extraction layer Stage2 of the backbone network with the high-level semantic features f5 output by the feature extraction layer Stage5 to extract the edge contour information of the camouflaged target, and the output of the edge extraction module EEM is normalized by the Sigmoid function to obtain a binary image f e , f e Upsample four times to get the edge prediction map O e , by disguising the target's edge label GE to O e For boundary supervision, the edge loss function used is L Dice .

4. The method for detecting disguised targets based on an attention mechanism and a convolutional neural network according to claim 1, wherein: A position embedding strategy is introduced in the position-aware recurrent convolution module PARCM.

5. The method for detecting camouflaged targets based on an attention mechanism and a convolutional neural network according to claim 1, wherein: In the channel attention component, nonlinear characteristics are introduced through the feedforward neural network FFN, and a channel attention mechanism SE Block is added after FFN to highlight key channels.

6. The method for detecting camouflaged targets based on an attention mechanism and a convolutional neural network according to claim 1, wherein: After the global features obtained in step S3 are effectively fused with the edge contour information obtained in step S4, multi-scale feature fusion is performed to obtain the multi-scale aggregated feature f 345 The specific contents include: using the edge enhancement module ERM to transform the edge contour information from the edge extraction module EEM into e Global features output by the position-aware recurrent convolution module PARCM Fusion is performed to obtain feature f i ',i∈{3,4,5}; Then, the multi-scale attention mechanism MSAM is introduced into the multi-scale fusion module MSFM to perform multi-scale feature fusion, where the feature f i ',i∈{4,5} is fused by multi-scale feature fusion module MSFM to obtain feature f 45 , then the feature f 45 The feature f3' is fused through the MSFM multi-scale fusion module to obtain the multi-scale aggregation feature f 345 .

7. The method for detecting camouflaged targets based on an attention mechanism and a convolutional neural network according to claim 6, wherein: The frequency channel attention component FcaNet is introduced into the edge enhancement module ERM, and the two-dimensional discrete cosine transform DCT is used to extract information of different frequency components and combine them.

8. The method for detecting disguised targets based on an attention mechanism and a convolutional neural network according to claim 1, wherein: In step S6, the structured loss function selected for deep supervision of the binary label map of the disguised target is the weighted binary cross loss and weighted intersection-over-union loss 9. The method for detecting disguised targets based on an attention mechanism and a convolutional neural network according to claim 1, wherein: The formula of the total loss function L of the disguised target detection network model is as follows: L=λ1L s +λ2L e L e =L Dice (O e ,GE) Among them, L s represents the structured loss, L e represents edge loss, λ1 and λ2 are weight factors of structural loss and edge loss respectively; s is the predicted camouflage target saliency map, O e Represents the predicted edge contour map of the camouflaged target; GT represents the saliency label of the camouflaged target, and GE represents the edge label of the camouflaged target; is the weighted binary cross loss in the structured loss, is the weighted intersection-over-union loss in the structured loss, L Dice is the edge loss obtained using the Dice coefficient.

Images