A camouflaged target detection method based on Gaussian attention

Abstract

The present invention discloses a camouflaged target detection method based on Gaussian attention, which relates to the field of image processing technology. First, the data set of the image to be predicted is passed through a multi-scale feature extraction network based on Gaussian attention to extract multi-scale features with more local structural information of the camouflaged target; secondly, a feature enhancement method based on a graph convolutional neural network is used to enhance the extracted features, so as to more effectively model the local structure of the camouflaged target; then, the enhanced features are input into a hierarchical feature pyramid module to obtain further fused features, so as to perform camouflaged target detection in the prediction module, and an improved loss function that optimizes the global structure is used to perform loss supervision on the training of the prediction module. This method can solve the problem that the traditional method cannot effectively distinguish between the background and the camouflaged target, and improve the detection accuracy of the camouflaged target.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and more particularly to a camouflage target detection method based on Gaussian attention. Background Art

[0002] Camouflaged target detection is an emerging visual detection task, which aims to accurately and efficiently detect camouflaged targets that are "perfectly" embedded in the surrounding environment. At the same time, the difficulty lies in how to distinguish extremely similar backgrounds from camouflaged targets.

[0003] Most of the current camouflaged target detection methods use convolutional neural networks to extract camouflaged target features, but convolutional neural networks lose structural information during feature extraction, and the actual receptive field of convolutional neural networks is much smaller than the theoretical receptive field. Therefore, the camouflaged target detection model based on convolutional neural networks usually cannot fully capture the global context information. In addition, the current camouflaged target detection methods do not take into account the local information and global structure of the camouflaged target.

[0004] The camouflaged target detection method based on Gaussian attention can avoid the above-mentioned problems caused by feature extraction by convolutional neural networks and the inability to take into account both the local and the global aspects.

[0005] Therefore, how to distinguish extremely similar backgrounds and disguised targets based on deep learning methods is an urgent problem that technicians in this field need to solve. Summary of the invention

[0006] In view of this, the present invention provides a camouflaged target detection method based on Gaussian attention, which can accurately distinguish the camouflaged target and the background when facing the camouflaged target detection task.

[0007] In order to achieve the above object, the present invention adopts the following technical solution:

[0008] A camouflaged target detection method based on Gaussian attention comprises the following steps:

[0009] S1. Input the dataset image into the multi-scale feature extraction network based on Gaussian attention to extract multi-scale features with more local structural information of the camouflaged target;

[0010] S2, input the extracted multi-scale features into the graph convolution feature enhancement network for feature enhancement to obtain enhanced features;

[0011] S3, inputting the enhanced features into the hierarchical feature fusion module, accumulating the imperceptible camouflaged target information, and obtaining the fused features;

[0012] S4, input the fusion features into the prediction module to obtain the prediction value;

[0013] S5, input the predicted value into the improved loss function calculation module, calculate the loss between the predicted value and the true value, and train the prediction module according to the loss;

[0014] S6: Obtain the image to be detected and input it into the trained prediction module to obtain the camouflaged target segmentation image.

[0015] Preferably, the multi-scale feature extraction network includes a block module, a linear embedding layer, a first level, a second level, a third level and a fourth level; the process of performing multi-scale feature extraction is:

[0016] S11: Input the dataset image into the block division module for block division to obtain image blocks; each 4x4 adjacent pixels is an image block, which is expanded in the direction of the dataset image channel;

[0017] S12: The image block performs linear transformation on the channel data of each pixel through the linear embedding layer to obtain a feature block;

[0018] S13: The feature block is input to the first level for Gaussian attention calculation to obtain the output feature

[0019] S14: Output features Input the second layer, perform downsampling and Gaussian attention calculation in turn to obtain output features

[0020] S15: Output features Input the third level, perform downsampling and Gaussian attention calculation in turn to obtain output features

[0021] S16: Output features Input the fourth level, perform downsampling and Gaussian attention calculation in turn to obtain output features

[0022] Preferably, the first level includes 2 Gaussian attention modules; each Gaussian attention module includes 2 layer normalization layers, 1 Gaussian attention layer and 1 fully connected layer; each Gaussian attention module performs Gaussian attention calculation on the feature block to obtain a set of output features The specific process is:

[0023] S131: feature block input layer normalization layer, obtain normalized features;

[0024] S132: Input the normalized features into the Gaussian attention layer, perform Gaussian attention calculation, and obtain the post-attention features;

[0025] S133: concatenating the post-attention feature and the feature block to obtain the first feature;

[0026] S134: The concatenated features are sequentially passed through a layer normalization layer and a fully connected layer to obtain a second feature;

[0027] S135: Concatenate the first feature and the second feature to obtain an output feature

[0028] Preferably, the specific process of performing Gaussian attention calculation on the normalized features in S132 is:

[0029] S1321: Based on the normalized features and three parameter matrices W q , W k , W v , calculate and obtain the three matrices Q, K, and V; the expression is:

[0030] Q=XW q , K = XW k ,V=XW v ;

[0031] Among them, X represents the normalized feature;

[0032] S1322: Calculate the Gaussian prior weight G of each spatial position in the normalized feature; the expression is:

[0033]

[0034] Among them, (x, y) represents the spatial position coordinates; σ represents the standard deviation; t represents the dimension of the normalized feature;

[0035] S1323: According to the three matrices Q, K, V and Gaussian prior weights, the post-attention features are obtained through softmax function processing;

[0036] The dot product operation is performed on the transpose of the matrix Q and the matrix K. The dot product operation result is divided by the set scaling factor and then added to the Gaussian prior weight. The addition result is processed by the softmax function and then tensor multiplied with the V matrix to obtain the attention feature Attention(Q, K, V); the expression is:

[0037]

[0038] Where d represents the dimension of each head in the multi-head attention mechanism.

[0039] Preferably, the second layer and the fourth layer include 1 downsampling module and 2 Gaussian attention modules respectively, and the third layer includes 6 Gaussian attention modules; the process of Gaussian attention calculation in the second layer, the third layer and the fourth layer is the same as that in the first layer.

[0040] Preferably, the specific process of enhancing the features in S2 is:

[0041] S21: Output features Output characteristics Output characteristics and output features Perform pairing, normalize the paired feature pairs, and obtain the feature pairs to be enhanced;

[0042] S22: each feature to be enhanced in each feature pair to be enhanced is processed by two linear projection functions to obtain a reduced dimension feature 1 and a reduced dimension feature 2, and the reduced dimension feature 1 and the reduced dimension feature 2 are concatenated to obtain a concatenated feature;

[0043] S23: The concatenated features are processed by a linear projection function, and then a softmax function is used to generate a weight map. The dimension reduction feature 1 is weighted by element-by-element multiplication according to the weight map, and an adaptive average pooling operation is performed on the weighted result to obtain the processed features;

[0044] S24: Perform matrix product operation on the processed features and the dimension reduction feature 1, and process the matrix product result using the softmax function to obtain the attention map T a ;

[0045] S25: Through the attention map T a The reduced dimension feature 2 is projected into the graph domain and then input into the graph convolutional neural network to output the enhanced features.

[0046] Preferably, the hierarchical feature fusion module includes several layers of aggregation modules, and the specific implementation process of S3 is:

[0047] S31: Aggregate the enhanced features corresponding to each feature pair to be enhanced to obtain an aggregated feature set;

[0048] S32: Divide the adjacent aggregated features in the aggregated feature set into a pair of aggregated feature pairs, and pass them to the aggregation module of the next layer;

[0049] S33: The aggregation module of the next layer aggregates each pair of aggregation features to obtain the aggregation feature set of the next layer, and returns to S32 until the aggregation module of the last layer completes the aggregation and obtains a fusion feature.

[0050] Preferably, the calculation formula for aggregation performed by the aggregation module is:

[0051] F p =CBR(Cat(CBR(Cat(F p-1 , F i )), F i-1 ))

[0052] F′ i =Up (CBR(F p ))

[0053] Among them, F i and F i-1 Represents the adjacent feature pairs of the current layer; F p-1 It is the aggregate feature output by the aggregation module of the previous layer; Cat represents the concatenation operation; CBR represents the convolution layer, batch normalization layer and activation function, that is, it is processed by the convolution layer, batch normalization layer and activation function in sequence.

[0054] Preferably, the improved loss function is expressed as:

[0055] L=L BCEL +λL UAL +L IoU

[0056]

[0057]

[0058]

[0059] Among them, l i，j represents the loss of each pixel; g i，j represents the ground truth label of the pixel; p i，j represents the predicted probability of a pixel as a camouflaged target; λ represents the set hyperparameter; the predicted value includes the predicted probability of each pixel.

[0060] Preferably, the prediction module adopts a neural network module, including 2 1x1 convolutional layers, a batch normalization layer, a ReLU activation function and a bilinear upsampling layer.

[0061] It can be seen from the above technical solution that compared with the prior art, the present invention discloses a camouflaged target detection method based on Gaussian attention. First, the data set image is input into a multi-scale feature extraction network based on Gaussian attention to extract multi-scale features with more local structural information of the camouflaged target; the extracted features are then input into a graph convolution feature enhancement network for feature enhancement; further, the enhanced features are input into a hierarchical feature fusion module to accumulate imperceptible camouflaged target information, and the fused features are input into a prediction module to obtain a predicted value; finally, the predicted value is input into an improved loss function calculation module to calculate the loss with the true value, the prediction module is trained using the loss, and the trained prediction module is used to perform camouflaged target detection on the image to be detected to obtain a camouflaged target segmentation image. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0063] Figure 1 A flow chart of a camouflaged target detection method based on Gaussian attention provided by the present invention;

[0064] Figure 2 A schematic diagram of a camouflaged target detection structure based on Gaussian attention provided by the present invention;

[0065] Figure 3 A schematic diagram of the Gaussian attention module structure provided by the present invention;

[0066] Figure 4 Schematic diagram of the Gaussian attention mechanism provided by the present invention. DETAILED DESCRIPTION

[0067] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0068] Example 1

[0069] The embodiment of the present invention discloses a camouflaged target detection method based on Gaussian attention. Figure 1 As shown, the following steps are included:

[0070] S1. Input the dataset image into the multi-scale feature extraction network based on Gaussian attention to extract multi-scale features with more local structural information of the camouflaged target;

[0071] S2, input the extracted multi-scale features into the graph convolution feature enhancement network for feature enhancement;

[0072] S3, inputting the enhanced multi-scale features into the hierarchical feature fusion module to accumulate the imperceptible camouflaged target information and obtain the fusion features;

[0073] S4, input the fusion features into the prediction module to obtain the prediction value;

[0074] S5, input the predicted value into the improved loss function calculation module, calculate the loss between the predicted value and the true value, and train the prediction module according to the loss;

[0075] S6: Obtain the image to be detected and input it into the trained prediction module to obtain the camouflaged target segmentation image.

[0076] The Gaussian attention-based camouflaged target detection method provided in this embodiment can accurately distinguish the camouflaged target from the background when faced with a detection task where it is difficult to distinguish between the background and the camouflaged target.

[0077] Example 2

[0078] In a specific embodiment, the camouflaged target detection method is described in detail:

[0079] S1: Input the dataset image into the multi-scale feature extraction network based on Gaussian attention. The multi-scale feature extraction network based on Gaussian attention includes a block module, a linear embedding layer, a first layer, a second layer, a third layer and a fourth layer. The first layer includes two Gaussian attention modules, the second layer includes a downsampling module and two Gaussian attention modules, the third layer includes a downsampling module and six Gaussian attention modules, and the fourth layer includes a downsampling module and two Gaussian attention modules.

[0080] Specifically, first, the RGB three-channel image with a width of H and a height of W is input into the block module for block division. Each 4×4 adjacent pixels is a block. After passing through the block module, the feature shape changes from (H, W, 3) to (H / 4, W / 4, 48);

[0081] Then, the linear embedding layer is used to perform a linear transformation on the channel data of each pixel from 48 to C, that is, the feature shape is changed from (H / 4, W / 4, 48) to (H / 4, W / 4, C), and the feature block Y is obtained;

[0082] Next, the features are passed through two Gaussian attention transformer modules in turn to construct the first level. Specifically, the feature block Y is input into the normalization module to obtain the normalized feature X, and the normalized feature X is input into Figure 3 The Gaussian attention module shown above obtains the features after attention

[0083] Reference Figure 4 As shown, the calculation process of the Gaussian attention module is:

[0084] First, three trainable parameter matrices W are used q , W k , W v To calculate the Q, K, and V matrices of feature X, the formula is as follows:

[0085] Q=XW q , K = XWk , V = XW v ;

[0086] Then calculate the Gaussian prior weight of each spatial position in feature X, the formula is as follows:

[0087]

[0088] In the above formula, σ is the standard deviation, (x, y) represents the position coordinates in feature X;

[0089] Dot product operation on the transpose of Q and K, then divide by a scaling factor and add the calculated Gaussian prior weight G, and then use the softmax function to process the result and finally perform tensor multiplication with V to obtain the post-attention feature. The formula is as follows:

[0090]

[0091] Post-attention features Concatenate with the unnormalized feature block Y to get the feature The features after stitching The input layer is normalized and then fed into the fully connected layer to obtain features. and with Concatenate to get output features

[0092] Will Down-sample to construct the second level (the third and fourth levels are the same), and pass through two Gaussian attention transformer modules in turn to obtain features Specifically, each 2x2 adjacent pixel is divided into a block in the downsampling module, and the obtained feature map is spliced ​​in the depth direction. A layer normalization module is passed, and then a fully connected layer is used to make a linear change in the depth direction of the feature map, changing the depth of the feature map from C to C / 2. After downsampling, the height and width of the feature map will be halved, and the depth will be doubled; then two Gaussian attention transformer modules are used to obtain the feature Among them, the downsampling module includes a block module, a splicing module, a layer normalization module and a fully connected layer;

[0093] Will Downsample, build the third layer, and pass through six Gaussian attention transformer modules in sequence to obtain features

[0094] Will Downsample, build the fourth level, and pass through two Gaussian attention transformer modules in turn to obtain features

[0095] S2: Extract multi-scale features There are 12 feature vectors in total, which are paired in pairs to obtain six pairs of features. Each feature is normalized and then enhanced. The multi-scale feature extraction network based on the Gaussian attention mechanism brings powerful global context modeling capabilities, but lacks a local mechanism for information exchange in local areas. The camouflaged target always shares very similar appearance information with the noise object and the background, and the slight differences are difficult to distinguish through low-order relationships. Therefore, a local feature enhancement module is designed to be applied to adjacent features (local areas) to enhance local feature representation.

[0096] Take one pair of features as an example, and the same goes for the other feature pairs:

[0097] After a pair of features are normalized, we get features T1 and T2. We concatenate features T1 and T2 to get concatenated feature T q , and pass both feature T1 and feature T2 through two linear projection functions ω v and ω k , get the dimension reduction feature T v With the dimension reduction feature T k ;

[0098] Taking feature T1 as an example, the same is true for feature T2. The formula is as follows:

[0099] T v =ω v (T1)

[0100] T k =ω k (T1):

[0101] The obtained splicing feature T q Through the linear projection function ω q , and then use the softmax function to generate the weight map And use this weight map to transform T by element-wise multiplication k Weighted, then adaptive average pooling Get the feature T' q , to reduce the computational cost, the formula is as follows:

[0102]

[0103] T' q With T k Perform matrix product and use the softmax function to generate the attention map T a , the formula is as follows:

[0104]

[0105] The attention map T a and T v Input to the graph fusion module, through the attention graph T a T v Projected into the graph domain, the formula is as follows:

[0106]

[0107] The feature T g Send it to the graph convolutional neural network ReLU and output the final feature after feature enhancement The formula is as follows:

[0108]

[0109] In the above formula, A is the adjacency matrix encoding the connectivity of the graph, ω g is the weight of the graph convolutional neural network;

[0110] The graph convolutional neural network ReLU is a neural network layer with the following layer-by-layer propagation rules:

[0111]

[0112] A is the connection matrix of the input graph, I is the identity matrix; for The degree matrix of each layer; H is the feature of each layer; σ is the nonlinear activation function; W is the trainable weight matrix of a specific layer;

[0113] S3: Features after feature enhancement A total of 12 feature vectors are input into the hierarchical feature fusion module, and the current adjacent enhanced feature pair and the aggregated features output by the previous aggregation module are interacted and merged through the aggregation module; if it is the first layer, it does not need to interact and merge with the aggregated features output by the previous aggregation module;

[0114] The current aggregated feature is passed to the next layer and the next aggregation module until it is aggregated into an aggregated feature. The aggregation module is calculated using the following formula:

[0115] F p =CBR(Cat(CBR(Cat(F p-1 , F i )), F i-1 ))

[0116] F' i =Up(CBR(F p ))

[0117] In the above formula, assuming that Fi and F i-1 is the adjacent feature pair of the current layer, F p-1 It is the feature aggregated by the previous aggregation module. Cat refers to the concatenation operation, and CBR refers to the convolution layer plus the batch normalization layer plus the activation function.

[0118] S4: Input the fused features into the prediction module to obtain the predicted value, and use the improved loss function formula to calculate the loss. The formula is as follows:

[0119] L=L BCEL +λL UAL +L IoU

[0120]

[0121]

[0122]

[0123] In the above formula, where l i，j is the loss of each pixel, g i，j is the ground truth label of the pixel, p i，j is the predicted probability of a pixel being a camouflaged target, and λ is a hyperparameter;

[0124] The prediction module is a neural network module, including a 1x1 convolution layer, a batch normalization layer, a ReLU activation function, a bilinear upsampling layer, and another 1x1 convolution layer, and finally outputs a tensor with a single channel, which is usually used for binary classification tasks.

[0125] S5: According to the loss training prediction module, the collected image to be detected is input into the trained prediction module, and the prediction value of the image to be detected is output, and the camouflaged target segmentation image is obtained according to the prediction value. According to the prediction probability (0-1) of each pixel point in the prediction value being a camouflaged target, the similar points closer to 1 are visually set to white, and finally the camouflaged target segmentation image is obtained.

[0126] The local information and global structural information of the disguised target are two important factors that affect the detection performance of the disguised target. In the traditional disguised target detection method based on deep learning, if only a single-scale feature extraction network is used, the underlying features lack semantic information, and the high-level features are prone to lose information, which seriously affects the detection performance of the disguised target. In this embodiment, a multi-scale feature extraction network based on Gaussian attention is used to extract multi-level image features and use Gaussian prior weighted attention weights, which can more effectively model the local structure of the disguised target. In addition, in the face of the extracted multi-scale features, if feature fusion or parallel strategies are directly adopted, the expression ability of the multi-scale features to the image will be weakened, thereby affecting the detection performance of the disguised target. In this embodiment, a feature enhancement method based on a graph convolutional neural network is first used to enhance the representation ability of multi-scale features. In the face of the enhanced features, this embodiment then adopts a hierarchical feature fusion method to further fuse the enhanced features and accumulate effective information that is difficult to detect in the disguised target. Finally, in the traditional camouflaged target detection method based on deep learning, a common binary cross entropy loss function is used for loss calculation, but the binary cross entropy loss independently calculates the loss of each pixel, ignoring the global structure of the image, which is unfriendly for the detection task of distinguishing extremely similar camouflaged targets and backgrounds. In this embodiment, an improved loss function is first used to optimize the global structure. The method of the present invention can effectively solve the problem that the traditional method cannot effectively distinguish the background from the camouflaged target, and improve the detection accuracy of the camouflaged target detection.

[0127] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0128] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A camouflaged target detection method based on Gaussian attention, characterized in that: The steps include: S1: Input the dataset image into the multi-scale feature extraction network based on Gaussian attention to extract multi-scale features; S2: Input the extracted multi-scale features into the graph convolutional feature enhancement network for feature enhancement to obtain enhanced features; S3: Input the enhanced features into the hierarchical feature fusion module to obtain fused features; S4: Input the fused features into the prediction module to obtain the predicted value; S5: input the predicted value into the improved loss function calculation module, calculate the loss between the predicted value and the true value, and train the prediction module according to the loss; S6: Obtain the image to be detected and input it into the trained prediction module to obtain the camouflaged target segmentation image; The multi-scale feature extraction network includes a block module, a linear embedding layer, a first layer, a second layer, a third layer, and a fourth layer; the process of multi-scale feature extraction is: S11: Input the dataset image into the block division module for block division to obtain image blocks; S12: The image block performs linear transformation on the channel data of each pixel through the linear embedding layer to obtain a feature block; S13: The feature block is input to the first level for Gaussian attention calculation to obtain the output feature S14: Output features Input the second layer, perform downsampling and Gaussian attention calculation in turn to obtain output features S15: Output features Input the third level, perform downsampling and Gaussian attention calculation in turn to obtain output features S16: Output features Input the fourth level, perform downsampling and Gaussian attention calculation in turn to obtain output features The first level includes 2 Gaussian attention modules; each Gaussian attention module includes 2 layer normalization layers, 1 Gaussian attention layer and 1 fully connected layer; each Gaussian attention module performs Gaussian attention calculation on the feature block to obtain a set of output features The specific process is: S131: feature block input layer normalization layer, obtain normalized features; S132: Input the normalized features into the Gaussian attention layer, perform Gaussian attention calculation, and obtain the post-attention features; S133: concatenating the post-attention feature and the feature block to obtain the first feature; S134: The concatenated features are sequentially passed through a layer normalization layer and a fully connected layer to obtain a second feature; S135: Concatenate the first feature and the second feature to obtain an output feature The specific process of performing Gaussian attention calculation on the normalized features in S132 is: S1321: Based on the normalized features and three parameter matrices W q , W k , W v , calculate and obtain the three matrices Q, K, and V; S1322: Calculate the Gaussian prior weight of each spatial position in the normalized feature; S1323: According to the three matrices Q, K, V and Gaussian prior weights, the post-attention features are obtained through softmax function processing; The specific process of enhancing features in S2 is: S21: Output features Output characteristics Output characteristics and output features Perform pairing, normalize the paired feature pairs, and obtain the feature pairs to be enhanced; S22: each feature to be enhanced in each feature pair to be enhanced is processed by two linear projection functions to obtain a reduced dimension feature 1 and a reduced dimension feature 2, and the reduced dimension feature 1 and the reduced dimension feature 2 are concatenated to obtain a concatenated feature; S23: The concatenated features are processed by a linear projection function, and then a softmax function is used to generate a weight map. The dimension reduction feature 1 is weighted by element-by-element multiplication according to the weight map, and an adaptive average pooling operation is performed on the weighted result to obtain the processed features; S24: Perform matrix product operation on the processed features and the dimension reduction feature 1, and process the matrix product result with the softmax function to obtain the attention map T a ; S25: Through the attention map T a Project the reduced dimension feature 2 into the graph domain and input it into the graph convolutional neural network to output the enhanced features; The improved loss function calculation module is expressed as: L=l BCEL +l UAL +l IoU Among them, l i，j represents the loss of each pixel; g i，j represents the ground truth label of the pixel; p i，j represents the predicted probability of a pixel as a camouflaged target; λ represents the set hyperparameter; the predicted value includes the predicted probability of each pixel.

2. The camouflaged target detection method based on Gaussian attention according to claim 1, characterized in that: The second and fourth layers include one downsampling module and two Gaussian attention modules respectively, and the third layer includes one downsampling module and six Gaussian attention modules; The process of Gaussian attention calculation at the second, third, and fourth levels is the same as that at the first level.

3. The camouflaged target detection method based on Gaussian attention according to claim 1, characterized in that: The hierarchical feature fusion module includes several layers of aggregation modules. The specific implementation process of S3 is as follows: S31: Aggregate the enhanced features corresponding to each feature pair to be enhanced to obtain an aggregated feature set; S32: Divide the adjacent aggregated features in the aggregated feature set into a pair of aggregated feature pairs, and pass them to the aggregation module of the next layer; S33: The aggregation module of the next layer aggregates each pair of aggregation features to obtain the aggregation feature set of the next layer, and returns to S32 until the aggregation module of the last layer completes the aggregation and obtains the fusion feature.

4. The camouflaged target detection method based on Gaussian attention according to claim 1, characterized in that: The prediction module adopts a neural network module, including 2 1×1 convolutional layers, a batch normalization layer, a ReLU activation function and a bilinear upsampling layer.

Images