A camouflaged target detection method based on the interaction of multi-scale context and multi-level features based on stereo attention

Abstract

The present invention discloses a camouflaged target detection method based on multi-scale context and multi-level feature interaction of stereo attention, which includes a backbone network, a multi-scale context exploration module (MCE), a multi-level feature interaction module (MFI) and a stereo attention enhancement module (SAE). The backbone network extracts multi-level features f of the input image. i (i=1,2,3,4); the Multi-scale Context Exploration Module (MCE) progressively captures strongly correlated multi-scale features through multi-receptive field convolution operations; the Multi-level Feature Interaction Module (MFI) adaptively aggregates multi-level features through a hybrid interaction strategy to generate powerful feature representations; and the Stereo Attention Enhancement Module (SAE) adaptively assigns weights to the aggregated features to filter out unimportant information. It is worth mentioning that this invention not only improves the performance of disguised object detection methods but also achieves significant progress in the field of salient object detection.

Description

Technical Field

[0001] The present invention relates to the fields of image processing and computer vision, and in particular to a camouflaged target detection method based on the interaction of multi-scale context and multi-level features of stereo attention. Background Art

[0002] Camouflaged object detection (COD) is an emerging visual inspection task that aims to mimic human visual detection mechanisms to accurately locate and completely segment objects that have similar texture, color, and pattern to the background within an input image. In recent years, camouflaged object detection has attracted increasing attention in the computer vision community due to its high application value in various fields, such as medical image segmentation, agricultural pest detection, and military camouflage object detection.

[0003] Early camouflaged object detection relied primarily on traditional handcrafted features to distinguish foreground from background, using image texture differences, color contrast, motion, intensity differences, optical flow, or multimodal fusion methods that combine these features. However, these methods only exploit low-level features. In complex scenarios, such as those with very similar foreground and background, or when the object's edges are blurred, these algorithms struggled to generate accurate predictions.

[0004] In recent years, with the advancement of deep learning, convolutional neural network (CNN)-based camouflaged object detection methods have broken through the limitations of hand-crafted features and achieved significant progress. Le et al. proposed a detection algorithm, ANet, with a multi-task learning strategy. This algorithm combines classification and segmentation tasks. The classification stream identifies whether a camouflaged object exists in an image, and if so, the segmentation stream performs segmentation. Inspired by the predator hunting process, Fan et al. proposed a two-stage camouflaged object detection algorithm based on a coarse-to-fine strategy: search followed by recognition. The search phase locates camouflaged objects in the image, while the recognition phase refines the detected objects. Furthermore, Fan et al. also proposed the largest dataset to date, COD10K, which contains 10,000 images across 78 classes. Pang et al. employed a scale-up / down strategy, feeding images of three different resolutions into a triplet network for scale integration and suppressing background interference through a designed uncertainty-aware loss. Lü et al. innovatively proposed two tasks: camouflaged object ranking and camouflaged object localization, to estimate the difficulty of camouflaged objects and identify their discriminative regions. Zhang et al. used existing monocular depth prediction methods to generate depth maps on a camouflaged target dataset, supplementing RGB images and achieving significant performance gains. Sun et al. believed that the two key factors in solving the COD problem are a large receptive field that provides rich context and cross-level feature fusion. Therefore, the authors proposed two corresponding modules to achieve better performance. Although the aforementioned camouflaged target detection methods recognize the importance of contextual information during feature extraction, they rarely consider the correlation between contextual information at different scales, resulting in inaccurate prediction maps. Therefore, a method is needed to explore the correlation between multi-scale context to improve detection accuracy. Summary of the Invention

[0005] To alleviate these issues, this paper proposes a disguised target detection method based on the interaction of multi-scale context and multi-level features using stereo attention. This method captures strongly correlated multi-scale features, adaptively aggregates multi-level features through a hybrid interaction strategy, and finally adaptively assigns feature weights through attention to generate a powerful feature representation. Experimental results demonstrate that this method significantly improves performance over existing methods.

[0006] The present invention discloses a camouflaged target detection method based on the interaction of multi-scale context and multi-level features of stereo attention, which comprises the following steps:

[0007] 1. Through the backbone network Swin-Transformer, the features of the input disguised target image and the salient target image are extracted respectively to obtain the initial features f i (i=1,2,3,4).

[0008] 1.1) Collect and organize datasets related to camouflaged object detection and salient object detection. The camouflaged object detection datasets include CHAMELEON, CAMO, COD10K, and NC4K; the salient object detection datasets include SOD, ECSSD, PASCAL-S, DUT-OMRON, HKU-IS, and DUST.

[0009] 1.2) Considering that model training requires large-scale data, for the camouflaged object detection task, this patent uses the CAMO-Train dataset and the COD10K-Train dataset, totaling 4,040 images containing camouflaged objects, as the training dataset, and uses CHAMELEON, CAMO-Test, COD10K-Test, and NC4K as the test datasets; for the salient object detection task, this patent uses the DUTS dataset, containing 10,553 images, as the training dataset, and uses the ECSSD, PASCAL-S, HKU-IS, DUT-OMRON, and DUTS-TE datasets as the test datasets.

[0010] 1.3) Using the Swin-Transformer pre-trained on ImageNet as the backbone network, the features of the input disguised target image and the salient target image are extracted to obtain the initial features f i (i=1,2,3,4);

[0011] 2. The initial feature f is explored through the multi-scale context exploration module (MCE) i (i=1,2,3,4) to perform feature enhancement to capture strongly correlated multi-scale features I i (i=1,2,3,4).

[0012] 2.1) To fully enhance the initial features f extracted by the backbone network i (i=1, 2, 3, 4), the multi-scale context exploration module (MCE) is divided into four branches, which gradually integrate the multi-scale features processed by MCE.

[0013] 2.2) Specifically, the initial feature f i (i=1, 2, 3, 4) first undergoes group convolution processing with a group number of 4 to reduce the number of channels and improve computational efficiency (initial feature f i(i=2, 3, 4) are added to the features of the previous MCE branch and then the following operations are performed. These features are then fed into a convolution operation C with a kernel size of 1×1×128 and an asymmetric convolution operation A with a kernel size of (1×(2k+1), (2k+1)×1), where k represents the kth branch of the MCE. Convolution operation C is used to reconstruct the initial features into 128 channels, and asymmetric convolution is used to extract contextual information. The convolved features are element-wise added to the input features to increase feature diversity, and then further extracted through a depthwise separable convolution with a padding rate of 2k. Compared with conventional convolution operations, depthwise separable convolution has a lower number of parameters and lower computational cost.

[0014] 2.3) As can be seen from the above description, the features extracted by the previous branch using a small receptive field are input to the next branch with a large receptive field as guidance, thereby enhancing the correlation between multi-scale contexts. The above process can be formulated as:

[0015]

[0016] Among them, down represents the dimensionality reduction operation, C represents the convolution operation with a convolution kernel size of 1×1, A represents the asymmetric convolution operation with a convolution kernel size of (1×(2k+1), (2k+1)×1), ∑ represents the element-by-element addition operation, and D k represents a depthwise separable convolution with a padding of 2k.

[0017] 2.4) S obtained from each branch of MCE k (k=1, 2, 3, 4) are cascaded and combined with the initial feature f i (i=1, 2, 3, 4) are added respectively, and finally a multi-scale feature I is generated. i (i=1, 2, 3, 4). The above process can be expressed as:

[0018] I=∑(down(f),down(Cat(S1,...,S k ))), k=4 (2)

[0019] Among them, down represents the dimensionality reduction operation, ∑ represents the element-by-element addition operation, and Cat represents the cascade operation.

[0020] 3. Adaptive multi-scale feature I through a hybrid interaction strategy using a multi-level feature interaction module (MFI) i (i=1, 2, 3, 4) are aggregated to generate multi-level features y i (i=1, 2, 3, 4).

[0021] 3.1) The MFI module mixes and interacts multiple levels of features in a top-down manner. High-level features interact with low-level features in turn. Multi-scale features I i (i=1, 2, 3, 4) act as guides to participate in the interaction of features at the same level.

[0022] 3.2) Taking the generation of multi-level feature y2 as an example, the multi-level features y3 and y4 generated by the previous MFI module and the multi-scale feature I2 of the corresponding level are first subjected to element-by-element multiplication and addition operations, followed by element-by-element addition and multiplication operations. Finally, the two sets of features are cascaded to obtain the multi-level feature y2. The generation of features at other levels is similar to that of y2. The above process can be expressed as:

[0023]

[0024] Among them, Cat represents the cascade operation, and ∑ and * represent element-by-element addition and multiplication operations.

[0025] 4. Adaptively assign weights to the aggregated multi-level features through the Stereo Attention Enhancement Module (SAE) to filter out unimportant information and ultimately obtain high-quality output features. i (i=1, 2, 3, 4).

[0026] 4.1) First, global average pooling and global maximum pooling operations are performed on the multi-level features aggregated by the previous MFI module to focus on global context information and reduce the interference of useless information; then, the features of these two branches are respectively subjected to k×k one-dimensional convolution, ReLU operation, and k×k one-dimensional convolution to enhance the feature representation, where the convolution kernel size k can be adaptively selected; finally, the two enhanced features are added together and then activated by Sigmoid to obtain the channel attention weight a i .

[0027] 4.2) In order to make full use of the multi-level features after aggregation, i (i=1, 2, 3, 4) performs a series of 2D convolution operations. Specifically, the operation is divided into two branches. The first branch performs 2D convolutions with kernel sizes of 1×1, k×k, and 1×1, and the second branch performs a 1×1 2D convolution. Note that each convolution operation includes BatchNorm and ReLU operations; then, the features of the two branches are added element by element, and after convolution and activation, they are combined with a i Multiply element by element to get the spatial attention weight b i .

[0028] 4.3) In order to retain the original features, the multiplied weights are connected with the multi-level feature residuals to finally generate a powerful feature representation; after dimensionality reduction, the output feature O is obtained i(i=1, 2, 3, 4). The above process can be expressed as:

[0029]

[0030] Among them, σ represents the Sigmoid function, C i Represents a convolution operation with a convolution kernel size of i×i, ReLU activation, avg global average pooling, max global maximum pooling, C i BR represents a convolutional block (including convolutional layer, BatchNorm and activation layer).

[0031] 5. For each high-quality output feature O i (i=1, 2, 3, 4) performs dimensionality reduction to generate a saliency prediction map or a disguised prediction map, and supervises it with the true value map. BCE loss and IoU loss are used as loss functions to supervise the training model. The total loss function can be expressed as:

[0032]

[0033] in, represents the binary cross entropy loss, which can calculate the loss of each predicted pixel independently; represents the IoU loss, which can increase the response of pixels while considering the global context.

[0034] The present invention has the following beneficial effects:

[0035] By adopting the above scheme, the present invention solves the problem that existing disguised target detection methods pay little attention to the correlation between contextual information of different scales. By capturing strongly correlated multi-scale features, aggregating multi-level features with a hybrid interaction strategy, and focusing on the main parts of the image, a high-quality feature representation is finally generated, thereby improving the performance of the disguised target detection method. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a network model structure diagram.

[0037] Figure 2 It is the structural diagram of the multi-scale context exploration module MCE.

[0038] Figure 3 It is the structural diagram of the multi-level feature interaction module MFI.

[0039] Figure 4 It is the stereo attention enhancement module SAE.

[0040] Figure 5 This is the experimental result of camouflaged target detection.

[0041] Figure 6 This is the result of the salient target detection experiment. DETAILED DESCRIPTION

[0042] The technical solutions in the embodiments of the present invention will be described clearly and completely below with reference to the accompanying drawings in the examples of the present invention.

[0043] The flowchart framework of the present invention is as follows Figure 1 As shown in Figure 1, a disguised target detection method based on the interaction of multi-scale context and multi-level features of stereo attention is proposed. The specific operation is described as follows:

[0044] 1. Through the backbone network Swin-Transformer, the features of the input disguised target image and the salient target image are extracted respectively to obtain the initial features f i (i=1,2,3,4).

[0045] 1.1) Collect and organize datasets related to camouflaged object detection and salient object detection. The camouflaged object detection datasets include CHAMELEON, CAMO, COD10K, and NC4K; the salient object detection datasets include SOD, ECSSD, PASCAL-S, DUT-OMRON, HKU-IS, and DUST.

[0046] 1.2) Considering that model training requires large-scale data, for the camouflaged object detection task, this patent uses the CAMO-Train dataset and the COD10K-Train dataset, totaling 4,040 images containing camouflaged objects, as the training dataset, and uses CHAMELEON, CAMO-Test, COD10K-Test, and NC4K as the test datasets; for the salient object detection task, this patent uses the DUTS dataset, containing 10,553 images, as the training dataset, and uses the ECSSD, PASCAL-S, HKU-IS, DUT-OMRON, and DUTS-TE datasets as the test datasets.

[0047] 1.3) Using the Swin-Transformer pre-trained on ImageNet as the backbone network, the features of the input disguised target image and the salient target image are extracted to obtain the initial features f i (i=1, 2, 3, 4);

[0048] 2. The initial feature f is explored through the multi-scale context exploration module (MCE) i (i=1, 2, 3, 4) to perform feature enhancement to capture strongly correlated multi-scale features I i (i=1,2,3,4), the detailed process is as follows Figure 2 shown.

[0049] 2.1) To fully enhance the initial features f extracted by the backbone network i (i=1, 2, 3, 4), the multi-scale context exploration module (MCE) is divided into four branches, which gradually integrate the multi-scale features processed by MCE.

[0050] 2.2) Specifically, the initial feature f i (i=1, 2, 3, 4) first undergoes group convolution processing with a group number of 4 to reduce the number of channels and improve computational efficiency (initial feature f i (i=2, 3, 4) are added to the features of the previous MCE branch and then the following operations are performed. These features are then fed into a convolution operation C with a kernel size of 1×1×128 and an asymmetric convolution operation A with a kernel size of (1×(2k+1), (2k+1)×1), where k represents the kth branch of the MCE. Convolution operation C is used to reconstruct the initial features into 128 channels, and asymmetric convolution is used to extract contextual information. The convolved features are element-wise added to the input features to increase feature diversity, and then pass through a depthwise separable convolution with a padding rate of 2k to further extract important multi-scale contextual information. Compared with conventional convolution operations, depthwise separable convolution has a lower number of parameters and lower computational cost.

[0051] 2.3) As can be seen from the above description, the features extracted by the previous branch using a small receptive field are input to the next branch with a large receptive field as guidance, thereby enhancing the correlation between multi-scale contexts. The above process can be formulated as:

[0052]

[0053] Among them, down represents the dimensionality reduction operation, C represents the convolution operation with a convolution kernel size of 1×1, A represents the asymmetric convolution operation with a convolution kernel size of (1×(2k+1), (2k+1)×1), ∑ represents the element-by-element addition operation, and D k represents a depthwise separable convolution with a padding of 2k.

[0054] 2.4) S obtained from each branch of MCE k (k=1, 2, 3, 4) are cascaded and combined with the initial feature f i (i=1, 2, 3, 4) are added respectively, and finally a multi-scale feature I is generated. i (i=1, 2, 3, 4). The above process can be expressed as:

[0055] I=∑(down(f),down(Cat(S1,...,S k ))), k=4 (2)

[0056] Among them, down represents the dimensionality reduction operation, ∑ represents the element-by-element addition operation, and Cat represents the cascade operation.

[0057] 3. Adaptive multi-scale feature I through a hybrid interaction strategy using a multi-level feature interaction module (MFI) i (i=1, 2, 3, 4) are aggregated to generate multi-level features y i (i=1,2,3,4), the specific process is as follows Figure 3 shown.

[0058] 3.1) The MFI module mixes and interacts multiple levels of features in a top-down manner. High-level features interact with low-level features in turn. Multi-scale features I i (i=1, 2, 3, 4) act as guides to participate in the interaction of features at the same level.

[0059] 3.2) Taking the generation of multi-level feature y2 as an example, the multi-level features y3 and y4 generated by the previous MFI module and the multi-scale feature I2 of the corresponding level are first subjected to element-by-element multiplication and addition operations, followed by element-by-element addition and multiplication operations. Finally, the two sets of features are cascaded to obtain the multi-level feature y2. The generation of features at other levels is similar to that of y2. The above process can be expressed as:

[0060]

[0061] Among them, Cat represents the cascade operation, and ∑ and * represent element-by-element addition and multiplication operations.

[0062] 4. Adaptively assign weights to the aggregated multi-level features through the Stereo Attention Enhancement Module (SAE) to filter out unimportant information and ultimately obtain high-quality output features. i (i=1,2,3,4), the specific process is as follows Figure 4 shown.

[0063] 4.1) First, global average pooling and global maximum pooling operations are performed on the multi-level features aggregated by the previous MFI module to focus on global context information and reduce the interference of useless information; then, the features of these two branches are respectively subjected to k×k one-dimensional convolution, ReLU operation, and k×k one-dimensional convolution to enhance the feature representation, where the convolution kernel size k can be adaptively selected; finally, the two enhanced features are added together and then activated by Sigmoid to obtain the channel attention weight a i .

[0064] 4.2) In order to make full use of the multi-level features after aggregation, i(i=1, 2, 3, 4) performs a series of 2D convolution operations. Specifically, the operation is divided into two branches. The first branch performs 2D convolutions with kernel sizes of 1×1, k×k, and 1×1, and the second branch performs a 1×1 2D convolution. Note that each convolution operation includes BatchNorm and ReLU operations; then, the features of the two branches are added element by element, and after convolution and activation, they are combined with a i Multiply element by element to get the spatial attention weight b i .

[0065] 4.3) In order to retain the original features, the multiplied weights are connected with the multi-level feature residuals to finally generate a powerful feature representation; after dimensionality reduction, the output feature O is obtained i (i=1, 2, 3, 4). The above process can be expressed as:

[0066]

[0067] Among them, σ represents the Sigmoid function, C i represents a convolution operation with a convolution kernel size of i×i, ReLU activation, avg global average pooling, max global maximum pooling, C i BR represents a convolutional block (including convolutional layer, BatchNorm and activation layer).

[0068] 5. For each high-quality output feature O i (i=1, 2, 3, 4) performs dimensionality reduction to generate a saliency prediction map or a disguised prediction map, and supervises it with the true value map. BCE loss and IoU loss are used as loss functions to supervise the training model. The total loss function can be expressed as:

[0069]

[0070] in, represents the binary cross entropy loss, which can calculate the loss of each predicted pixel independently; represents the IoU loss, which can increase the response of pixels while considering the global context.

[0071] We compared the prediction graph generated based on the above algorithm and supervised training strategy with the existing methods. The results of disguised target detection are shown in the figure below. Figure 5 As shown in the figure, the results of salient target detection are as follows Figure 6 shown.

[0072] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. Any other changes or other equivalent replacements that do not deviate from the technical solution of the present invention are included in the scope of protection of the present invention.

Claims

1. A camouflaged target detection method based on the interaction of multi-scale context and multi-level features based on stereo attention, characterized by: The following steps are involved: 1) The backbone network Swin-Transformer is used to extract features of the input salient target image and the disguised target image respectively to obtain the initial features f i (i=1,2,3,4); 2) The initial feature f is explored through the multi-scale context exploration module MCE i (i=1,2,3,4) to perform feature enhancement to capture strongly correlated multi-scale features I i (i=1,2,3,4); 2.1) To fully enhance the initial features f extracted by the backbone network i (i=1, 2, 3, 4), the multi-scale context exploration module MCE is divided into four branches, which gradually integrate the multi-scale features processed by MCE; 2.2) Specifically, the initial feature f i (i=1,2,3,4) first undergoes group convolution processing with a group number of 4 to reduce the number of channels and improve computational efficiency, where the initial feature f i After adding the features of the previous branch of the MCE (i=2,3,4), the following operations are performed. These features are then used as inputs and passed in parallel through a convolution operation C with a convolution kernel size of 1×1×128 and an asymmetric convolution operation A with a convolution kernel size of (1×(2k+1),(2k+1)×1), where k represents the kth branch of the MCE. The convolution operation C is used to reconstruct the initial features into 128 channels, and the asymmetric convolution is used to extract contextual information. The convolved features are element-wise added to the input features to increase feature diversity, and then passed through a depthwise separable convolution with a padding rate of 2k to further extract important multi-scale contextual information. Compared with conventional convolution operations, depthwise separable convolution has a lower number of parameters and lower computational cost. 2.3) From the above description, it can be seen that the features extracted by the previous branch using a small receptive field are input to the next branch with a large receptive field as guidance, thereby enhancing the correlation between multi-scale contexts; The above process can be formulated as: Among them, down represents the dimensionality reduction operation, C represents the convolution operation with a convolution kernel size of 1×1, A represents the asymmetric convolution operation with a convolution kernel size of (1×(2k+1), (2k+1)×1), ∑ represents the element-by-element addition operation, and D k represents a depth-wise separable convolution with a filling rate of 2k; 2.4) S obtained from each branch of MCE k (k=1,2,3,4) are cascaded and combined with the initial feature f i (i=1,2,3,4) are added respectively, and finally the multi-scale feature I is generated i (i=1,2,3,4); the above process can be expressed as: I=∑(donw(f),down(Cat(S1,…,S k ))),k=4(2) Among them, down represents the dimensionality reduction operation, ∑ represents the element-by-element addition operation, and Cat represents the cascade operation; 3) Adaptive multi-scale feature I through the multi-level feature interaction module MFI with a hybrid interaction strategy i (i=1,2,3,4) are aggregated to generate multi-level features y i (i=1,2,3,4); 4) The stereo attention enhancement module SAE is used to adaptively assign weights to the aggregated multi-level features to filter out unimportant information and ultimately obtain high-quality output features. i (i=1,2,3,4); 5) For each high-quality output feature O i (i=1,2,3,4) performs dimensionality reduction to generate saliency prediction maps or disguised prediction maps, and supervises them with the true value maps.

2. The method for detecting disguised targets based on multi-scale context and multi-level feature interaction based on stereo attention according to claim 1, characterized in that: The specific method of step 1) is: 2.1) Collect and organize datasets related to camouflage object detection and salient object detection. The camouflage object detection datasets include CHAMELEON, CAMO, COD10K, and NC4K; the salient object detection datasets include SOD, ECSSD, PASCAL-S, DUT-OMRON, HKU-IS, and DUST. 2.2) Considering that model training requires large amounts of data, for the camouflaged object detection task, we use the CAMO-Train and COD10K-Train datasets, totaling 4,040 images containing camouflaged objects, as training datasets, and CHAMELEON, CAMO-Test, COD10K-Test, and NC4K as test datasets. For the salient object detection task, we use the DUTS dataset, containing 10,553 images, as training datasets, and the ECSSD, PASCAL-S, HKU-IS, DUT-OMRON, and DUTS-TE datasets as test datasets. 2.3) Using the Swin-Transformer pre-trained on ImageNet as the backbone network, the features of the input disguised target image and the salient target image are extracted to obtain the initial features f i (i=1,2,3,4).

3. The method for detecting disguised targets based on multi-scale context and multi-level feature interaction based on stereo attention according to claim 1, characterized in that: The specific method of step 3) is: 3.1) The MFI module mixes and interacts multiple levels of features in a top-down manner. High-level features interact with low-level features in turn. Multi-scale features I i (i=1,2,3,4) as a guide to participate in the interaction of the same-level features; 3.2) Taking the generation of multi-level feature y2 as an example, the multi-level features y3 and y4 generated by the previous MFI module and the multi-scale feature I2 of the corresponding level are first subjected to element-by-element multiplication and element-by-element addition operations, and then element-by-element addition and element-by-element multiplication operations. Finally, the two sets of features are cascaded to obtain the multi-level feature y2. The generation of features of other levels is similar to that of y2. The above process can be expressed as: Among them, Cat represents the cascade operation, and ∑ and * represent element-by-element addition and multiplication operations.

4. The method for detecting disguised targets based on multi-scale context and multi-level feature interaction based on stereo attention according to claim 1, characterized in that: The specific method of step 4) is: 4.1) First, perform global average pooling and global maximum pooling operations on the multi-level features aggregated by the previous MFI module to focus on global context information and reduce the interference of useless information; Then, the features of the two branches are further enhanced by k×k one-dimensional convolution, ReLU operation, and k×k one-dimensional convolution, where the convolution kernel size k can be adaptively selected; finally, the two enhanced features are added together and then activated by Sigmoid to obtain the channel attention weight a i ; 4.2) In order to make full use of the multi-level features after aggregation, i (i=1,2,3,4) performs a series of two-dimensional convolution operations; specifically, the operation is divided into two branches. The first branch performs two-dimensional convolution with kernel sizes of 1×1, k×k, and 1×1 in sequence, and the second branch performs a 1×1 two-dimensional convolution, where each convolution operation contains BatchNorm and ReLU operations; then, the features of the two branches are added element by element, and after convolution and activation, they are combined with a i Multiply element by element to get the spatial attention weight b i ; 4.3) In order to retain the original features, the multiplied weights are connected with the multi-level feature residuals to finally generate a powerful feature representation; after dimensionality reduction, the output feature O is obtained i (i=1,2,3,4); the above process can be expressed as: Among them, σ represents the Sigmoid function, C i represents a convolution operation with a convolution kernel size of i×i, ReLU activation, avg global average pooling, max global maximum pooling, C i BR represents a convolutional block, which includes a convolutional layer, BatchNorm and an activation layer.

5. The method for detecting disguised targets based on multi-scale context and multi-level feature interaction based on stereo attention according to claim 1, characterized in that: The specific method of step 5) is: use BCE loss and IoU loss as loss functions to supervise the training model. The total loss function can be expressed as: in, represents the binary cross entropy loss, which can calculate the loss of each predicted pixel independently; represents the IoU loss, which can increase the response of pixels while considering the global context.

Images