CN116703950A - Camouflage target image segmentation method and system based on multi-level feature fusion - Google Patents

Abstract

The invention discloses a camouflage target image segmentation method and a camouflage target image segmentation system based on multi-level feature fusion, wherein the method carries out global feature enhancement on a first feature image of each level and carries out local feature enhancement on a second feature image of each level; carrying out feature fusion on the reinforced local features of each level and the reinforced global features of the same level to obtain fusion features of multiple levels; conducting boundary guiding on the fusion characteristics of two shallow network layers in the fusion characteristics of multiple layers to obtain a boundary map; performing feature interaction on fusion features of adjacent network layers in the fusion features of multiple layers to obtain multiple interaction features; respectively carrying out boundary fusion on the boundary map and each interactive feature in the plurality of interactive features to obtain a plurality of boundary fusion features; based on the plurality of boundary fusion features, a camouflage target image in the camouflage target images to be segmented corresponding to each boundary fusion feature is segmented. The invention can improve the accuracy of the camouflage target image segmentation.

Description

A Camouflage Target Image Segmentation Method and System Based on Multi-level Feature Fusion

technical field

The invention relates to the technical field of camouflage target image segmentation, in particular to a camouflage target image segmentation method and system based on multi-level feature fusion.

Background technique

Camouflaged Object Segmentation (COS) aims to segment camouflaged objects with high similarity to the background. COS uses a computer vision model to assist the human visual and perceptual system for image segmentation of camouflaged targets.

However, in the existing technology, the model with CNN (such as ResNet) as the backbone has a strong ability to extract local features, while CNN has limited ability to obtain long-range feature dependencies due to the limitation of the receptive field; the model with transformer (such as visiontransformer) as the backbone Benefiting from the attention mechanism in the transformer, it has a strong ability to model global feature relationships, but it has limitations in capturing fine-grained details, resulting in a weakened ability to express local features. However, the segmentation of camouflage target image needs to separate the target from the whole based on global features, and also needs to process detailed information such as boundaries based on local features. Using a single master network requires complex methods to fuse local features and global features, which is less efficient. . Most methods use simple operations to fuse multi-level features, such as concatenation and addition. When high-level features interact with low-level features, the two features are first fused by an addition operation. Then, the fused features are sent to the Sigmoid activation function to obtain a normalized feature map, and the normalized feature map is regarded as a feature-level attention map to enhance the feature representation. In this case, using the fused feature maps obtained by the simple addition operation to achieve cross-level feature enhancement cannot capture valuable information that is highly relevant to segmenting camouflage targets. Some models focus on extracting global texture features of camouflaged objects, ignoring the impact of boundaries on model expressiveness, and these models perform poorly when the target object shares the same texture features with the background. Since the texture of most camouflaged objects is similar to the background, distinguishing subtle differences in boundary local information is particularly important to improve model performance. Although some models consider boundary features, they often supervise the predicted boundary map as an independent branch without other processing, and the boundary map information is not fully utilized.

In summary, the existing techniques are difficult to capture practical feature information, and the predicted boundary map information cannot be fully utilized, therefore, it is difficult to achieve accurate camouflage target segmentation.

Contents of the invention

The present invention aims to solve at least one of the technical problems existing in the prior art. For this reason, the present invention proposes a method and system for segmenting a camouflaged target image based on multi-level feature fusion, which can improve the accuracy of segmenting a camouflaged target image.

In the first aspect, an embodiment of the present invention provides a method for segmenting a camouflage target image based on multi-level feature fusion, wherein the segmentation of a camouflage target image based on multi-level feature fusion includes:

Obtain the camouflaged target image to be segmented;

The first branch network and the second branch network use different network layers to perform multi-level feature extraction on the camouflaged target image to be segmented, and obtain the first feature map of various levels output by the first branch network and the second feature map. Multiple levels of second feature maps output by the two-branch network;

Perform global feature enhancement on the first feature map of each level to obtain multi-level enhanced global features; perform local feature enhancement on the second feature map of each level to obtain multi-level enhanced local features features; and performing feature fusion of the enhanced local features of each level and the enhanced global features of the same level to obtain fusion features of multiple levels;

performing boundary guidance on the fusion features of the two shallow network layers in the fusion features of the multiple levels to obtain a boundary map;

performing feature interaction on the fusion features of adjacent network layers among the fusion features of the multiple levels to obtain multiple interaction features;

performing boundary fusion on the boundary map and each of the plurality of interaction features, to obtain a plurality of boundary fusion features;

Based on the plurality of boundary fusion features, a camouflage target image in the to-be-segmented camouflage target images corresponding to each of the boundary fusion features is segmented.

Compared with the prior art, the first aspect of the present invention has the following beneficial effects:

This method uses different network layers to perform multi-level feature extraction on the camouflage target image to be segmented through the first branch network and the second branch network, and obtains the first feature map of various levels output by the first branch network and the output of the second branch network. The second feature map of multiple levels can better extract the features in the target image; the global feature enhancement is performed on the first feature map of each level to obtain the global features after multi-level enhancement, and the second feature map of each level Local feature enhancement is performed on the two feature maps to obtain enhanced local features of various levels, and the enhanced local features of each level are fused with the enhanced global features of the same level to obtain fusion features of multiple levels , by enhancing local features and global features, and fusing the enhanced local features and global features, the local features and global features can complement each other, providing comprehensive feature information for accurate segmentation of camouflaged target images; for multiple levels In the fusion features of the two shallow network layers, the fusion features of the two shallow network layers are used for boundary guidance to obtain a boundary map. Since more semantic information is retained in the shallow layer, the shallow fusion features are used for boundary guidance, which can generate high-quality boundaries. Fig. The fusion features of the adjacent network layers in the multi-level fusion features are interacted to obtain multiple interactive features, and the fusion features of multiple levels can complement each other to obtain a comprehensive feature expression; the boundary map is respectively combined with multiple Boundary fusion is performed on each interaction feature in the interaction feature to obtain multiple boundary fusion features. Based on the multiple boundary fusion features, the camouflage target image in the camouflage target image to be segmented corresponding to each boundary fusion feature is segmented, and the boundary map Guided by the boundary information in , the features of the boundary map are integrated with the interaction features of different levels, and the boundary features are refined to ensure the clarity and integrity of the boundary, which helps to distinguish the fine foreground and background boundaries of the camouflage target, and the Segmentation is more expressive and improves the accuracy of segmentation of camouflaged target images.

According to some embodiments of the present invention, the first branch network and the second branch network use different network layers to perform multi-level feature extraction on the to-be-segmented masquerade target image, and obtain various types of output from the first branch network. The first feature map of the level and the second feature map of various levels output by the second branch network include:

Using different network layers to perform feature extraction on the global context information of the camouflage target image to be segmented through the first branch network, and obtain first feature maps of various levels output by the first branch network;

The second branch network uses different network layers to perform feature extraction on the local detail information of the to-be-segmented masquerade target image, and obtain second feature maps of various levels output by the second branch network.

According to some embodiments of the present invention, the global feature enhancement is performed on the first feature map of each level to obtain global features after multi-level enhancement; the local feature is performed on the second feature map of each level Enhancing, obtaining enhanced local features of multiple levels; and performing feature fusion of the enhanced local features of each level and the enhanced global features of the same level to obtain fusion features of multiple levels, including :

Using a residual channel attention mechanism to perform global feature enhancement on the first feature map of each level to obtain global features after multiple level enhancements;

Using a spatial attention mechanism to enhance the local features of the second feature map of each level to obtain local features enhanced at multiple levels;

Concatenate the enhanced local features of each level with the enhanced global features of the same level to obtain multiple levels of spliced features, and use convolutional layers to promote the multiple levels of spliced features Feature fusion to obtain multiple levels of fusion features.

According to some embodiments of the present invention, performing boundary guidance on the fusion features of two shallow network layers among the fusion features of the various levels to obtain a boundary map includes:

Convolving the fusion features of the two shallow network layers in the fusion features of the various levels to obtain the first convolution feature and the second convolution feature;

performing an addition operation on the first convolutional feature and the second convolutional feature to obtain an added feature, and performing boundary guidance on the added feature by using multiple convolutional layers to obtain a boundary map.

According to some embodiments of the present invention, performing feature interaction on the fusion features of adjacent network layers among the fusion features of multiple levels to obtain multiple interaction features, including:

Introduce a multi-scale channel attention mechanism in the feature interaction module, and add the fusion features of adjacent network layers among the fusion features of the multiple levels to obtain multiple addition features;

Each of the added features is input to the multi-scale channel attention mechanism to obtain multiple multi-scale channel features;

Using an activation function for each of the multi-scale channel features to obtain a plurality of normalized features, and subtracting each of the normalized features by one to obtain a plurality of normalized difference features;

performing feature enhancement on the plurality of normalized features and the plurality of normalized difference features to obtain a plurality of enhanced normalized features and a plurality of enhanced normalized difference features;

Residually connect each of the enhanced normalized features and its corresponding fusion features to obtain multiple first residual features, and convolve each of the first residual features to obtain multiple first convolution feature;

Residually connect each of the enhanced normalized difference features and its corresponding fusion features to obtain a plurality of second residual features, and perform convolution on each of the second residual features to obtain a plurality of second convolutional features;

Adding each of the first convolution features and the corresponding second convolution features to obtain a plurality of added convolution features, and using the convolution layer to promote the plurality of added convolution features Fusion, multiple interactive features are obtained.

According to some embodiments of the present invention, the boundary fusion of the boundary map and each of the multiple interaction features to obtain multiple boundary fusion features includes:

Based on each of the interaction features, the target attention head branch is used to learn the overall feature of the target; wherein the target attention head branch is used to separate the target and the background as a whole based on the interaction features;

Based on the boundary map and each of the interaction features, use the boundary attention head branch to learn boundary detail features; wherein the boundary attention head branch is used to capture the sparseness of the target based on the boundary map and the interaction features Local boundary information;

The output of each of the target attention head branches and the output of each of the corresponding boundary attention head branches are spliced to obtain a plurality of output splicing features, and the multiple output splicing features are used in a convolutional layer Promote feature fusion and obtain multiple convolution fusion features;

A residual connection is performed between each convolutional fusion feature and each of the corresponding interaction features to obtain multiple boundary fusion features.

According to some embodiments of the present invention, the segmentation of the camouflage target image in the camouflage target image to be segmented corresponding to each of the boundary fusion features based on the plurality of boundary fusion features includes:

Inputting the plurality of boundary fusion features into a convolutional layer with a Sigmoid activation function generates a plurality of prediction maps;

Based on each of the prediction images, a masquerading target image among the masquerading target images to be segmented is segmented.

In the second aspect, the embodiment of the present invention also provides a multi-level feature fusion based camouflage target image segmentation system, the multi-level feature fusion based camouflage target image segmentation system includes:

A data acquisition unit, configured to acquire a camouflage target image to be segmented;

The feature extraction unit is used to perform multi-level feature extraction on the camouflage target image to be segmented by using different network layers through the first branch network and the second branch network, and obtain the first multi-level output of the first branch network. a feature map and multiple levels of second feature maps output by the second branch network;

The feature fusion unit is used to enhance the global features of the first feature map of each level to obtain global features after multiple level enhancements; perform local feature enhancement to the second feature map of each level to obtain multiple The enhanced local features of each level; and feature fusion of the enhanced local features of each level and the enhanced global features of the same level to obtain fusion features of multiple levels;

The boundary guidance unit is used to perform boundary guidance on the fusion features of the two shallow network layers in the fusion features of the various levels to obtain a boundary map;

A feature interaction unit, configured to perform feature interaction on the fusion features of adjacent network layers among the fusion features of the multiple levels to obtain multiple interaction features;

a boundary fusion unit, configured to perform boundary fusion on the boundary map and each of the plurality of interactive features to obtain multiple boundary fusion features;

An image segmentation unit, configured to segment a camouflage target image in the camouflage target images to be segmented corresponding to each of the boundary fusion features based on the plurality of boundary fusion features.

In a third aspect, an embodiment of the present invention also provides an electronic device, including:

at least one memory;

at least one processor;

at least one computer program;

The at least one computer program is stored in the at least one memory, and the at least one processor executes the at least one computer program to implement the multi-level feature fusion-based masquerade target image segmentation described in the first aspect above method.

In a fourth aspect, an embodiment of the present invention also provides a storage medium, the storage medium is a computer-readable storage medium, and the computer-readable storage medium stores a computer program, and the computer program is used to enable the computer to execute the above-mentioned first In one aspect, a camouflage target image segmentation method based on multi-level feature fusion is described.

It can be understood that the beneficial effects of the above-mentioned second aspect to the fourth aspect compared with the related technology are the same as those of the above-mentioned first aspect compared with the related technology. Please refer to the relevant description in the above-mentioned first aspect. This will not be repeated here.

Description of drawings

The above and/or additional aspects and advantages of the present invention will become apparent and understandable from the description of the embodiments in conjunction with the following drawings, wherein:

Fig. 1 is a flow chart of a method for segmenting a camouflaged target image based on multi-level feature fusion according to an embodiment of the present invention;

2 is a flow chart of a method for segmenting a camouflaged target image based on multi-level feature fusion according to another embodiment of the present invention;

Fig. 3 is the overall structure schematic diagram of the model of an embodiment of the present invention;

4 is a schematic diagram of a Res2Net module and a basic convolution module according to an embodiment of the present invention;

FIG. 5 is a structural diagram of a residual channel attention mechanism according to an embodiment of the present invention;

6 is a structural diagram of a spatial attention mechanism according to an embodiment of the present invention;

FIG. 7 is a structural diagram of an LGA module according to an embodiment of the present invention;

Fig. 8 is a structural diagram of MS-CA according to an embodiment of the present invention;

Fig. 9 is a structural diagram of a CFT module according to an embodiment of the present invention;

FIG. 10 is a structural diagram of an MTA according to an embodiment of the present invention;

Fig. 11 is a structural diagram of a BMTA according to an embodiment of the present invention;

Fig. 12 is a structural diagram of a BAH according to an embodiment of the present invention;

13 is a structural diagram of a camouflage target image segmentation system based on multi-level feature fusion according to an embodiment of the present invention;

Fig. 14 is a schematic diagram of a hardware structure of an electronic device according to an embodiment of the present invention.

Detailed ways

Embodiments of the present invention are described in detail below, examples of which are shown in the drawings, wherein the same or similar reference numerals designate the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by referring to the figures are exemplary only for explaining the present invention and should not be construed as limiting the present invention.

In the description of the present invention, if the first, second, etc. are described only for the purpose of distinguishing technical features, it cannot be understood as indicating or implying relative importance or implicitly indicating the number of indicated technical features or implying Indicates the sequence of the indicated technical features.

In the description of the present invention, it should be understood that when it comes to orientation descriptions, for example, the orientation or positional relationship indicated by up, down, etc. is based on the orientation or positional relationship shown in the drawings, which is only for the convenience of describing the present invention and simplifying the description , rather than indicating or implying that the device or element referred to must have a particular orientation, be constructed and operate in a particular orientation, and thus should not be construed as limiting the invention.

In the description of the present invention, it should be noted that, unless otherwise clearly defined, words such as setting, installation, and connection should be understood in a broad sense, and those skilled in the art can reasonably determine that the above words are included in the present invention in combination with the specific content of the technical solution. specific meaning.

In the existing technology, the model with CNN (such as ResNet) as the backbone has a strong ability to extract local features, while CNN has limited ability to obtain long-range feature dependencies due to the limitation of the receptive field; the model with transformer (such as visiontransformer) as the backbone benefits The attention mechanism in the transformer has a strong modeling ability for global feature relationships, but it has limitations in capturing fine-grained details, resulting in a weakened ability to express local features. Most methods use simple operations to fuse multi-level features, such as concatenation and addition. When high-level features interact with low-level features, the two features are first fused by an addition operation. Then, the fused features are sent to the Sigmoid activation function to obtain a normalized feature map, and the normalized feature map is regarded as a feature-level attention map to enhance the feature representation. In this case, using the fused feature maps obtained by the simple addition operation to achieve cross-level feature enhancement cannot capture valuable information that is highly relevant to segmenting camouflage targets. Some models focus on extracting global texture features of camouflaged objects, ignoring the impact of boundaries on model expressiveness, and these models perform poorly when the target object shares the same texture features with the background. Since the texture of most camouflaged objects is similar to the background, distinguishing subtle differences in boundary local information is particularly important to improve model performance. Although some models consider boundary features, they often supervise the predicted boundary map as an independent branch without other processing, and the boundary map information is not fully utilized.

In summary, the existing technology is difficult to capture practical feature information, so that the predicted boundary map information cannot be fully utilized, therefore, it is difficult to achieve accurate camouflaged target segmentation.

In order to solve the above problems, the present invention uses different network layers to perform multi-level feature extraction on the camouflage target image to be segmented through the first branch network and the second branch network, and obtains the first feature map and the second feature map of various levels output by the first branch network. The multi-level second feature map output by the two-branch network can better extract the features in the target image; the global feature enhancement is performed on the first feature map of each level, and the global features after multi-level enhancement are obtained. The second feature map of each level is enhanced with local features to obtain enhanced local features of multiple levels, and the enhanced local features of each level are fused with the enhanced global features of the same level to obtain multiple A level of fusion features, by enhancing local features and global features, and fusing the enhanced local features and global features, the local features and global features can complement each other and provide comprehensive feature information for accurate segmentation of camouflaged target images ;Boundary guidance is performed on the fusion features of two shallow network layers in the fusion features of multiple levels to obtain a boundary map. Since more semantic information is retained in the shallow layer, the boundary guidance is performed using shallow fusion features, which can Generate a high-quality boundary map; perform feature interaction on the fusion features of adjacent network layers in multiple levels of fusion features to obtain multiple interactive features, and the fusion features of multiple levels can complement each other to obtain a comprehensive feature expression; the boundary Boundary fusion is performed on the graph with each interaction feature in the multiple interaction features to obtain multiple boundary fusion features, and based on the multiple boundary fusion features, the camouflage target in the camouflage target image to be segmented corresponding to each boundary fusion feature is segmented. Image, guided by the boundary information in the boundary map, integrates the features of the boundary map with different levels of interaction features, refines the boundary features, ensures the clarity and integrity of the boundary, and helps to distinguish the fine foreground and background boundaries of camouflaged targets , the segmentation of the camouflage target has better expressiveness, and improves the accuracy of the camouflage target image segmentation.

Referring to FIG. 1 , an embodiment of the present invention provides a method for segmenting a camouflaged target image based on multi-level feature fusion. The segmentation of a camouflaged target image based on multi-level feature fusion includes but is not limited to steps S100 to S700, wherein:

Step S100, acquiring the camouflage target image to be segmented;

Step S200, use different network layers to perform multi-level feature extraction on the camouflage target image to be segmented through the first branch network and the second branch network, and obtain the first feature map of various levels output by the first branch network and the output of the second branch network The second feature map of multiple levels;

Step S300, perform global feature enhancement on the first feature map of each level to obtain multi-level enhanced global features; perform local feature enhancement on the second feature map of each level to obtain multi-level enhanced local features ; and the enhanced local features of each level are fused with the enhanced global features of the same level to obtain multi-level fusion features;

Step S400, perform boundary guidance on the fusion features of two shallow network layers among the fusion features of multiple levels, and obtain a boundary map;

Step S500, performing feature interaction on the fusion features of adjacent network layers among the fusion features of multiple levels to obtain multiple interaction features;

Step S600, performing boundary fusion on the boundary map and each of the multiple interaction features respectively, to obtain multiple boundary fusion features;

Step S700 , based on a plurality of boundary fusion features, segment a masquerade target image among the to-be-segmented masquerade target images corresponding to each boundary fusion feature.

In some embodiments from step S100 to step S700, in order to better extract the features in the target image, this embodiment uses different network layers through the first branch network and the second branch network to perform multi-level features on the masquerading target image to be segmented Extracting, obtaining the first feature map of multiple levels output by the first branch network and the second feature map of multiple levels output by the second branch network; in order to provide comprehensive feature information for accurate segmentation and camouflage target images, this embodiment By performing global feature enhancement on the first feature map of each level, global features after multi-level enhancement are obtained; local feature enhancement is performed on the second feature map of each level, and local features after multi-level enhancement are obtained; and The enhanced local features of each level are fused with the enhanced global features of the same level to obtain fusion features of multiple levels; in order to generate high-quality boundary maps, this embodiment uses multiple levels of The fusion features of the two shallow network layers in the fusion feature are guided by the boundary to obtain a boundary map; in order to obtain a comprehensive feature expression, this embodiment performs feature interaction by combining the fusion features of the adjacent network layers in the fusion features of multiple levels, Obtain multiple interactive features; in order to improve the accuracy of the segmentation of the camouflaged target image, this embodiment obtains multiple boundary fusion features by performing boundary fusion on the boundary map with each of the multiple interactive features, based on multiple boundary The features are fused, and the camouflage target images in the camouflage target images to be segmented corresponding to each boundary fusion feature are segmented.

In some embodiments, the first branch network and the second branch network use different network layers to perform multi-level feature extraction on the camouflage target image to be segmented, and obtain the first feature map and the second feature map of various levels output by the first branch network. Multiple levels of second feature maps output by the branch network, including:

performing feature extraction on the global context information of the camouflage target image to be segmented by using different network layers through the first branch network, and obtaining first feature maps of various levels output by the first branch network;

The second branch network uses different network layers to perform feature extraction on the local detail information of the to-be-segmented camouflage target image, and obtains second feature maps of multiple levels output by the second branch network.

Specifically, Swin-TransformerV2 (ie, the first branch network) uses different network layers to perform feature extraction on the global context information of the camouflage target image to be segmented, and obtain the first feature map of multiple levels output by the first branch network; Swin- The multi-head self-attention mechanism in TransformerV2 can break through the limitation of the receptive field in CNN, model the context relationship pixel by pixel in the global scope, assign greater weight to important features, and make feature expression more abundant. Through Res2Net (that is, the second branch network), different network layers are used to extract the local detail information of the camouflaged target image to be segmented, and the second feature map of multiple levels output by the second branch network is obtained; Res2Net is stronger and more effective. The multi-level feature extraction capability can refine features at a finer-grained level, highlighting the difference between foreground and background.

In this embodiment, since most current models extract features based on a single backbone network, it is necessary to rely on complex methods to fuse local features and global features, and the efficiency is low. Therefore, in this embodiment, the first branch network and the second branch network use different network layers to perform multi-level feature extraction on the camouflage target image to be segmented, so that the features in the target image can be better extracted.

It should be noted that this embodiment uses Swin-TransformerV2 and Res2Net for feature extraction, but is not limited to Swin-TransformerV2 and Res2Net, and this embodiment can be modified according to actual conditions without specific limitations.

In some embodiments, global feature enhancement is performed on the first feature map of each level to obtain global features after multi-level enhancement; local feature enhancement is performed on the second feature map of each level to obtain multi-level enhanced features The local features of each level; and the enhanced local features of each level are fused with the enhanced global features of the same level to obtain multiple levels of fusion features, including:

The residual channel attention mechanism is used to enhance the global features of the first feature map of each level, and obtain the global features after multiple levels of enhancement;

The spatial attention mechanism is used to enhance the local features of the second feature map of each level, and obtain local features after multi-level enhancement;

Concatenate the enhanced local features of each level with the enhanced global features of the same level to obtain multiple levels of spliced features, and use convolutional layers to promote feature fusion for multiple levels of spliced features to obtain multiple Hierarchical fusion features.

Specifically, the local spatial details and global context information fusion module is used to enhance and fuse features. The local spatial details and global context information fusion module applies the channel attention mechanism and the spatial attention mechanism at the same time, and uses the residual channel attention mechanism to enhance the global features of the first feature map of each level, and obtain the global features after multiple levels of enhancement. features; use the spatial attention mechanism to enhance the local features of the second feature map of each level, and obtain the local features enhanced by multiple levels; the enhanced local features of each level and the enhanced global features of the same level The features are spliced to obtain multiple levels of spliced features, and the convolutional layer is used to promote feature fusion for multiple levels of spliced features to obtain multiple levels of fusion features.

In this embodiment, the fusion module of local spatial details and global context information simultaneously considers the global context and local details of the image to identify the overall trend of the image, which effectively complements the feature extraction capabilities of the two main branch networks. Global information is used to extract important global features that roughly estimate the position of the target object, and local information is used to extract fine-grained features of the object. Local features and global features complement each other to provide comprehensive feature information for accurate camouflaged target segmentation.

In some embodiments, boundary guidance is performed on the fusion features of two shallow network layers in the fusion features of various levels to obtain a boundary map, including:

Convolving the fusion features of two shallow network layers in the fusion features of multiple levels to obtain the first convolution feature and the second convolution feature;

The addition operation is performed on the first convolution feature and the second convolution feature to obtain the addition feature, and a plurality of convolution layers are used for boundary guidance on the addition feature to obtain a boundary map.

Specifically, the boundary guidance module is used to convolve the fusion features of two shallow network layers among the fusion features of various levels to obtain the first convolution feature and the second convolution feature; for the first convolution feature and the second The convolutional features are added to obtain the added features, and multiple convolutional layers are used for boundary guidance on the added features to obtain a boundary map.

In this embodiment, since more semantic information is retained in the shallow layer, the fusion features of the shallow layer are used for boundary guidance, and a high-quality boundary map can be generated.

In some embodiments, the fusion features of adjacent network layers among the fusion features of various levels are subjected to feature interaction to obtain multiple interaction features, including:

Introduce the multi-scale channel attention mechanism in the feature interaction module, and add the fusion features of adjacent network layers among the fusion features of multiple levels to obtain multiple addition features;

Input each added feature to the multi-scale channel attention mechanism to obtain multiple multi-scale channel features;

Use an activation function to obtain multiple normalized features for each multi-scale channel feature, and subtract each normalized feature by one to obtain multiple normalized difference features;

Performing feature enhancement on multiple normalized features and multiple normalized difference features to obtain multiple enhanced normalized features and multiple enhanced normalized difference features;

Residually connect each enhanced normalized feature and its corresponding fusion feature to obtain multiple first residual features, and convolve each first residual feature to obtain multiple first convolutions feature;

Residually connect each enhanced normalized difference feature and its corresponding fusion feature to obtain multiple second residual features, and perform convolution on each second residual feature to obtain multiple second convolution feature;

Add each first convolution feature and its corresponding second convolution feature to obtain multiple added convolution features, and use the convolution layer to promote fusion of multiple added convolution features to obtain multiple interactions feature.

In this embodiment, the feature interaction module introduces a multi-scale channel attention mechanism in order to achieve efficient interaction of cross-layer features and to deal with changes in target size in camouflaged target segmentation. The multi-scale channel attention mechanism has strong adaptability to different scale targets. The multi-scale channel attention mechanism is based on a dual-branch structure. One branch uses global average pooling to obtain global features and allocate more attention to large-scale objects. The other branch uses point convolution to obtain fine-grained local details, which is more conducive to Capture features of small objects. Different from other multi-scale attention mechanisms, the multi-scale channel attention mechanism uses point convolution in two branches to compress and restore the features of the channel dimension, thereby aggregating multi-scale channel information at different levels and effectively characterizing the multiple layers of the convolutional layer. scale information. The fusion features of multiple levels can complement each other to obtain a more comprehensive feature expression.

In some embodiments, boundary fusion is performed between the boundary map and each of the multiple interaction features to obtain multiple boundary fusion features, including:

Based on each interaction feature, the target attention head branch is used to learn the overall feature of the target; where the target attention head branch is used to separate the target and background from the whole based on the interaction feature;

Based on the boundary map and each interaction feature, the boundary attention head branch is used to learn the boundary detail features; where the boundary attention head branch is used to capture the sparse local boundary information of the target based on the boundary map and interaction features;

The output of each target attention head branch is spliced with the output of each corresponding boundary attention head branch to obtain multiple output splicing features, and the multiple output splicing features are used to promote feature fusion by convolutional layers, and multiple output splicing features are obtained. Convolution fusion features;

Residual connections are made between each convolutional fusion feature and each corresponding interaction feature to obtain multiple boundary fusion features.

In this embodiment, guided by the boundary information in the boundary map, the features of the boundary map are integrated with the interaction features of different levels, and the boundary features are refined to ensure the clarity and completeness of the boundary, which is helpful to distinguish the subtleties of camouflaged targets. Foreground and background borders.

In some embodiments, based on a plurality of boundary fusion features, the camouflage target image in the camouflage target image to be segmented corresponding to each boundary fusion feature is segmented, including:

Input multiple boundary fusion features into a convolutional layer with a Sigmoid activation function to generate multiple prediction maps;

Based on each prediction map, a masquerade target image among the masquerade target images to be segmented is segmented.

In this embodiment, based on multiple boundary fusion features, the segmentation of the camouflage target has better expressiveness, and the accuracy of the segmentation of the camouflage target image can be improved.

For the convenience of those skilled in the art to understand, a group of best embodiments are provided below:

Referring to Figures 2 to 3, the overall structure of the model in this embodiment includes feature extraction, local spatial details and global context information fusion module (LGA module) for fusing local features and global features, and feature interaction for cross-layer feature interaction module (CFT module), a boundary-guided module (BGM module) for predicting boundary maps, a boundary-guided multi-convolution head-transposed attention module (BMTA module) for fusing boundary features, and a prediction layer to generate the final segmentation map. First, the camouflage target image to be segmented is input to the parallel structure of Swin-TransformerV2 and Res2Net to extract multi-level features, and then the features with the same resolution in the two branches are sent to the LGA module to aggregate global information and local information, and then The features output by the LGA module of the adjacent network layer are interactively fused and enhanced through the CFT module, and the features a1 and a2 output by the LGA module corresponding to the previous shallow network layer are used as the input of the BGM module to generate a predicted boundary map (BM), and then The feature and boundary map output by the CFT module are sent to the BMTA module, where the boundary information and global information are fused, and finally sent to the prediction layer to generate a prediction map. Camouflage the target image. The prediction graphs P1, P2, and P3 in Fig. 2 to Fig. 3 are from rough to fine, P1 is the roughest, P3 is the clearest and most complete, and this embodiment takes P3 as the final result. P1, P2, P3, and BM are all supervised by the loss function, which guides the model to optimize parameters and improve segmentation accuracy. The specific steps are:

1. Feature extraction.

Using Swin-TransformerV2 and Res2Net dual-branch structure to extract multiple levels of features from the camouflage target image to be segmented. The multi-head self-attention mechanism in Swin-TransformerV2 can break through the limitation of the receptive field in CNN, model the context relationship pixel by pixel in the global scope, assign greater weight to important features, and make feature expression more abundant. Res2Net has stronger and more effective multi-level feature extraction capabilities, refines features at a finer-grained level, and highlights the difference between foreground and background. Unlike representing multi-level features in a hierarchical manner, Res2Net replaces the 3×3 convolutional layers with a series of convolutional groups. The comparison between the Res2Net module and the basic convolution module is shown in Figure 4. A in Figure 4 is the basic convolution module, and b in Figure 4 is the Res2Net module. After each group is processed by the 3×3 convolutional layer, it will have an output with a larger receptive field than itself. This grouping strategy can better process the feature map, and the larger the split level dimension is, the learning The richer the feature information is.

For a given image I∈R ^3×H×W , where H and W represent the height and width of the input image respectively, and 3 represents an RGB image, both Swin-TransformerV2 and Res2Net contain four stages, and the images are sent to Swin- TransformerV2 and Res2Net to generate multi-level feature maps from four stages. The feature map T _i (i=1,2,3,4) is generated through the Swin-TransformerV2 branch. The resolution of T ₁ is H/4×W/4, and the resolution of T ₄ is H/32×W/32. The feature map R _i (i=1, 2, 3, 4) is generated through the Res2Net branch. The resolution of R ₁ is H/4×W/4, and the resolution of R ₄ is H/32×W/32, that is, two The feature maps generated at the same stage of each branch have the same spatial size (that is, have the same resolution).

2. Fusion of local spatial details and global context information.

The feature maps obtained by Swin-TransformerV2 and Res2Net are input into the LGA module. The LGA module simultaneously considers the global context and local details of the image to identify the general trend of the image, effectively supplementing the feature extraction capabilities of the two main branch networks. Global information is used to extract important global features that roughly estimate the position of the target object, and local information is used to extract fine-grained features of the object. Local features and global features complement each other to provide comprehensive feature information for accurate camouflaged object segmentation (COS).

The LGA module applies both residual channel attention and spatial attention. The structure of the residual channel attention mechanism is shown in Fig. 5. For the input feature map X∈R ^C×H×W , where C, H, and W represent the number of channels, height, and width respectively, the residual channel attention mechanism first obtains a 1×1×C feature map through global average pooling, Obtain global important feature information, then use downsampling (implemented by 1×1 convolution) to compress the number of channels, and then restore to the original number of channels C by upsampling (implemented by 1×1 convolution), and obtain the Weight coefficient, multiplying the weight coefficient with the original feature X can get a more discriminative feature map. Residual channel attention assigns larger weights to important channels, thereby enhancing global features along the channel dimension.

The structure of the spatial attention mechanism is shown in Fig. 6. For the input feature map X∈R ^C×H×W , the spatial attention mechanism obtains the feature map F _max ∈ R ^1×H×W and F _avg by performing maximum pooling and average pooling on the channel dimension of the compressed feature. ∈R ^1×H×W , then concatenate F _max and F _avg , and use the convolution operation on the spliced feature map, and then use the Sigmoid activation function to generate a spatial attention map F _s ∈R ^1×H×W , will Multiplying the spatial attention map F _s with the input feature map X can assign larger weights to important spaces, thereby enhancing the local details of the spatial domain.

The structure of the LGA module is shown in Figure 7. Features with the same resolution (such as T ₁ and R ₁ ) extracted from the two branches of CNN (Res2Net is a type of CNN) and Transformer (Swin-TransformerV2 is a variant of Transformer) are sent to the LGA module, from The feature Fc of CNN is sent to the spatial attention (SA) branch to further enhance the local features extracted by CNN and suppress irrelevant regions; the feature Ft from Transformer is sent to the residual channel attention (RCA) branch to enhance Transformer extraction Global context features.

In order to make the RCA branch focus on the learning of global important features, Ft is residually connected with the RCA features. For the SA branch, in order to reduce the calculation amount of the model, the convolution operation is first performed to reduce the channel dimension, and then the convolution result is residually connected with the SA features, so that the SA branch can focus on learning spatial features. Then the output of the two branches is concatenated to integrate global position information and local detail information, and a 3×3 convolutional layer is used to promote feature fusion, so as to adaptively integrate local features and global dependencies.

3. Boundary map generation.

Shallow feature layers (such as T ₁ and T ₂ , R ₁ and R ₂ ) retain more edge spatial information of the target, while deep convolutional layers (such as T ₃ and T ₄ , R ₃ and R ₄ ) retains more semantic information, so shallow features (a1 and a2 in Figure 3) are used as input to the BGM module to generate a boundary map (BM), and the BM and the input image have the same space through upsampling size, and measure the resulting boundary map by the following binary cross-entropy loss function.

in, Indicates the ground truth value of the boundary map of the i-th image, /> represents the predicted boundary map for the i-th image.

4. Cross-layer feature interaction.

The CFT module introduces the Multi-Scale Channel Attention (MS-CA) mechanism to achieve efficient interaction of cross-layer features and cope with the variation of object size in COS. The MS-CA mechanism has strong adaptability to targets of different scales. The structure of the MS-CA mechanism is shown in Figure 8. The MS-CA mechanism is based on a dual-branch structure. One branch uses global average pooling to obtain global features and allocate more attention to large-scale objects. The other branch uses point convolution to obtain fine-grained local details, which is more conducive to capturing small characteristics of the object. Different from other multi-scale attention mechanisms, the MS-CA mechanism uses point convolution in two branches to compress and restore the features of the channel dimension, thereby aggregating the multi-scale channel information at different levels and effectively characterizing the multi-scale information of the convolutional layer. .

The overall structure of the feature interaction module is shown in Figure 9. The features a _h and a _l of adjacent layers are first added, and then sent to MS-CA to obtain multi-scale channel information, and then the Sigmoid activation function is used to obtain a normalized feature map F _s , F _s , 1-F _s (corresponding to the dotted arrows in Figure 9) are multiplied with a _l and a _h respectively to enhance the feature representation. In order to preserve the original information of each feature, the original feature and the enhanced feature are residually connected, and then the two branches are merged by adding, and a 3×3 convolutional layer is used to promote fusion, and the output F of the CFT module is obtained. _e .

The feature interaction module aggregates cross-layer image features with different levels and receptive fields by introducing the MS-CA mechanism to provide rich multi-level context information, and multi-level features interact to produce more effective and discriminative image information, Enables the model to adaptively segment objects of different sizes.

5. Boundary-guided multi-convolution head transposed attention.

The BMTA module effectively combines the local details of the predicted boundary map and the global information of the target in a multi-head attention manner. The BMTA module is based on the multi-convolution head transpose attention (MTA) to realize the interaction between local and non-local pixels. The structure of multi-convolution transpose attention is shown in Figure 10.

First, the input feature map is normalized (layer normalization), and then sent to three 1×1 convolutions and 3×3 depth convolutions to generate query (Q), key (K), value ( V), then the transposition of Q and the transposition of K perform matrix multiplication to generate an attention map (AM), and then perform matrix multiplication between AM and the transposition matrix of V to generate a new feature map. The calculation formula is as follows:

The structure of the BMTA module is shown in Figure 11. The input feature maps are respectively fed into the Boundary Attention Head (BAH) branch and Object Attention Head (OAH) branch to learn different features. The BAH branch is integrated into the boundary map to provide boundary priors, so that the model can better learn edge detail features, and OAH learns the overall features of the target.

The structure and calculation formula of OAH are consistent with MTA, and realize the global feature extraction across channels to separate the target and background as a whole.

BAH introduces the predicted binary boundary map (BM) into MTA to learn a boundary-enhanced representation to efficiently capture the important sparse local boundary information of objects. The structure of BAH is shown in Figure 12.

Specifically, Q and K obtained after the convolution operation are multiplied by BM to obtain Q _b and K _b , and then Q _b and K _b are multiplied to obtain an attention matrix with boundary information. The calculation formula of BAH is as follows:

V is a matrix without boundary information. In this way, features can be refined by establishing a pair relationship at the boundary to ensure the clarity and integrity of the boundary.

The BMTA module finally concatenates the outputs of OAH and BAH and sends them to a 3×3 convolution to realize the fusion of boundary and overall features. In order to avoid feature degradation, a residual connection is used to add the original and fused features. Guided by the boundary information, BMTA integrates the features of the boundary map with different levels of feature representations, which helps the model distinguish fine foreground and background boundaries of camouflaged objects.

6. Prediction map generation.

The features of BMTA are sent to a 3×3 convolutional layer with Sigmoid to generate a prediction map, and the prediction map P _i (i=1,2,3) of each layer is used by the BCE loss function and the IOU loss function. supervision to optimize the parameters of the entire model. The model of this embodiment adopts multiple supervision, so that the model can better express the segmentation of the camouflaged target. The overall loss function of the model is as follows:

in, Indicates the forecast map, /> represents the true value of the image, /> Denotes the i-th predicted map.

Referring to Fig. 13, the embodiment of the present invention also provides a camouflage target image segmentation system based on multi-level feature fusion, the camouflage target image segmentation system based on multi-level feature fusion includes a data acquisition unit 100, a feature extraction unit 200, a feature fusion Unit 300, boundary guidance unit 400, feature interaction unit 500, boundary fusion unit 600 and image segmentation unit 700, wherein:

A data acquisition unit 100, configured to acquire a masquerading target image to be segmented;

The feature extraction unit 200 is configured to perform multi-level feature extraction on the camouflage target image to be segmented by using different network layers through the first branch network and the second branch network, and obtain the first feature map and the second feature map of various levels output by the first branch network. Multiple levels of second feature maps output by the two-branch network;

The feature fusion unit 300 is configured to perform global feature enhancement on the first feature map of each level to obtain multi-level enhanced global features; perform local feature enhancement to the second feature map of each level to obtain multiple level enhancements The enhanced local features; and the enhanced local features of each level are fused with the enhanced global features of the same level to obtain multi-level fusion features;

The boundary guidance unit 400 is used to perform boundary guidance on the fusion features of two shallow network layers in the fusion features of multiple levels to obtain a boundary map;

The feature interaction unit 500 is used to perform feature interaction on the fusion features of adjacent network layers in the fusion features of multiple levels to obtain multiple interaction features;

A boundary fusion unit 600, configured to perform boundary fusion on the boundary map and each of the multiple interactive features respectively to obtain multiple boundary fusion features;

The image segmentation unit 700 is configured to segment a camouflage target image in the camouflage target images to be segmented corresponding to each boundary fusion feature based on a plurality of boundary fusion features.

It should be noted that since the multi-level feature fusion-based camouflage target image segmentation system in this embodiment is based on the same inventive concept as the above-mentioned multi-level feature fusion-based camouflage target image segmentation method, the method implementation The corresponding content in the example is also applicable to this system embodiment, and will not be described in detail here.

The embodiment of the present application also provides an electronic device, the electronic device includes: at least one memory, at least one processor, at least one computer program, at least one computer program is stored in at least one memory, and at least one processor executes at least one A computer program to implement any one of the multi-level feature fusion-based segmentation methods for camouflaged target images in the above embodiments. The electronic device may be any intelligent terminal including a tablet computer, a vehicle-mounted computer, and the like.

Referring to FIG. 14, FIG. 14 illustrates a hardware structure of an electronic device according to another embodiment, and the electronic device includes:

The processor 810 may be implemented by a general-purpose CPU (Central Processing Unit, central processing unit), a microprocessor, an application-specific integrated circuit (Application Specific Integrated Circuit, ASIC), or one or more integrated circuits, and is used to execute related programs, so as to realize The technical solutions provided by the embodiments of the present application;

The memory 820 may be implemented in the form of a read-only memory (ReadOnlyMemory, ROM), a static storage device, a dynamic storage device, or a random access memory (RandomAccessMemory, RAM). The memory 820 can store operating systems and other application programs. When implementing the technical solutions provided by the embodiments of this specification through software or firmware, the relevant program codes are stored in the memory 820 and called by the processor 810 to execute the implementation of the present application. An example of a camouflage target image segmentation method based on multi-level feature fusion;

The input/output interface 830 is used to realize information input and output;

The communication interface 840 is used to realize the communication and interaction between this device and other devices, which can realize communication through wired methods (such as USB, network cable, etc.) or wireless methods (such as mobile network, WIFI, Bluetooth, etc.);

bus 850, to transfer information between various components of the device (such as processor 810, memory 820, input/output interface 830, and communication interface 840);

The processor 810 , the memory 820 , the input/output interface 830 and the communication interface 840 are connected to each other within the device through the bus 850 .

The embodiment of the present application also provides a storage medium, the storage medium is a computer-readable storage medium, and the computer-readable storage medium stores a computer program, and the computer program is used to make the computer execute any one of the above-mentioned embodiments based on multi-level Feature Fusion for Image Segmentation of Camouflaged Objects.

As a non-transitory computer-readable storage medium, memory can be used to store non-transitory software programs and non-transitory computer-executable programs. In addition, the memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage, flash memory, or other non-transitory solid-state storage. In some embodiments, the memory optionally includes memory located remotely from the processor, and these remote memories may be connected to the processor via a network. Examples of the aforementioned networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

The embodiments described in the embodiments of the present application are to illustrate the technical solutions of the embodiments of the present application more clearly, and do not constitute a limitation to the technical solutions provided by the embodiments of the present application. Those skilled in the art know that with the evolution of technology and new For the emergence of application scenarios, the technical solutions provided by the embodiments of the present application are also applicable to similar technical problems.

Those skilled in the art can understand that the technical solution shown in Figure 1 does not constitute a limitation to the embodiment of the present application, and may include more or fewer steps than those shown in the illustration, or combine some steps, or different steps .

The device embodiments described above are only illustrative, and the units described as separate components may or may not be physically separated, that is, they may be located in one place, or may be distributed to multiple network units. Part or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

Those of ordinary skill in the art can understand that all or some of the steps in the methods disclosed above, the functional modules/units in the system, and the device can be implemented as software, firmware, hardware, and an appropriate combination thereof.

The terms "first", "second", "third", "fourth", etc. (if any) in the description of the present application and the above drawings are used to distinguish similar objects and not necessarily to describe specific sequence or sequence. It is to be understood that the data so used are interchangeable under appropriate circumstances such that the embodiments of the application described herein can be practiced in sequences other than those illustrated or described herein. Furthermore, the terms "comprising" and "having", as well as any variations thereof, are intended to cover a non-exclusive inclusion, for example, a process, method, system, product or device comprising a sequence of steps or elements is not necessarily limited to the expressly listed instead, may include other steps or elements not explicitly listed or inherent to the process, method, product or apparatus.

It should be understood that in this application, "at least one (item)" means one or more, and "multiple" means two or more. "And/or" is used to describe the association relationship of associated objects, indicating that there can be three types of relationships, for example, "A and/or B" can mean: only A exists, only B exists, and A and B exist at the same time , where A and B can be singular or plural. The character "/" generally indicates that the contextual objects are an "or" relationship. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one item (piece) of a, b or c can mean: a, b, c, "a and b", "a and c", "b and c", or "a and b and c ", where a, b, c can be single or multiple.

In the several embodiments provided in this application, it should be understood that the disclosed devices and methods may be implemented in other ways. For example, the device embodiments described above are only illustrative. For example, the division of the above units is only a logical function division. In actual implementation, there may be other division methods. For example, multiple units or components can be combined or can be Integrate into another system, or some features may be ignored, or not implemented. In another point, the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces, and the indirect coupling or communication connection of devices or units may be in electrical, mechanical or other forms.

The units described above as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or may be distributed to multiple network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, each unit may exist separately physically, or two or more units may be integrated into one unit. The above-mentioned integrated units can be implemented in the form of hardware or in the form of software functional units.

If the integrated unit is realized in the form of a software function unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or part of the contribution to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium , including multiple instructions to make a computer device (which may be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the method in each embodiment of the present application. The aforementioned storage media include: U disk, mobile hard disk, read-only memory (Read-Only Memory, ROM for short), random access memory (Random Access Memory, RAM for short), magnetic disk or optical disk, etc., which can store programs. medium.

The preferred embodiments of the embodiments of the present application have been described above with reference to the accompanying drawings, which does not limit the scope of rights of the embodiments of the present application. Any modifications, equivalent replacements and improvements made by those skilled in the art without departing from the scope and essence of the embodiments of the present application shall fall within the scope of rights of the embodiments of the present application.

Claims (10)

1. The camouflage target image segmentation method based on the multi-level feature fusion is characterized by comprising the following steps of:

obtaining a camouflage target image to be segmented;

performing multi-level feature extraction on the camouflage target image to be segmented by adopting different network layers through a first branch network and a second branch network to obtain a first feature image of multiple levels output by the first branch network and a second feature image of multiple levels output by the second branch network;

performing global feature enhancement on the first feature map of each level to obtain global features after multiple levels of enhancement; carrying out local feature enhancement on the second feature map of each level to obtain local features after multiple levels of enhancement; carrying out feature fusion on the enhanced local features of each level and the enhanced global features of the same level to obtain fusion features of multiple levels;

conducting boundary guiding on the fusion characteristics of two shallow network layers in the fusion characteristics of the multiple layers to obtain a boundary diagram;

performing feature interaction on the fusion features of adjacent network layers in the fusion features of the multiple layers to obtain multiple interaction features;

Respectively carrying out boundary fusion on the boundary map and each interactive feature in the plurality of interactive features to obtain a plurality of boundary fusion features;

and dividing a camouflage target image in the camouflage target image to be divided, which corresponds to each boundary fusion feature, based on the boundary fusion features.

2. The method for splitting a camouflage target image based on multi-level feature fusion according to claim 1, wherein the steps of extracting multi-level features of the camouflage target image to be split by using different network layers through a first branch network and a second branch network to obtain a first feature map of multiple layers output by the first branch network and a second feature map of multiple layers output by the second branch network comprise:

extracting features of global context information of the camouflage target image to be segmented by adopting different network layers through a first branch network to obtain a first feature map of multiple layers output by the first branch network;

and extracting the characteristics of the local detail information of the camouflage target image to be segmented by adopting different network layers through a second branch network, and obtaining a plurality of layers of second characteristic images output by the second branch network.

3. The method for dividing a camouflage target image based on multi-level feature fusion according to claim 1, wherein global feature enhancement is performed on the first feature map of each level to obtain global features with enhanced multiple levels; carrying out local feature enhancement on the second feature map of each level to obtain local features after multiple levels of enhancement; and performing feature fusion on the enhanced local feature of each level and the enhanced global feature of the same level to obtain fusion features of multiple levels, wherein the feature fusion comprises the following steps:

carrying out global feature enhancement on the first feature map of each level by adopting a residual channel attention mechanism to obtain global features after multiple levels of enhancement;

local feature enhancement is carried out on the second feature map of each level by adopting a spatial attention mechanism, so that local features after multiple levels of enhancement are obtained;

splicing the enhanced local features of each level with the enhanced global features of the same level to obtain splicing features of multiple levels, and adopting a convolution layer to promote feature fusion of the splicing features of the multiple levels to obtain fusion features of multiple levels.

4. The method for dividing a camouflage target image based on multi-level feature fusion according to claim 1, wherein the performing boundary guiding on the fusion features of two shallow network layers in the multi-level fusion features to obtain a boundary map comprises:

convolving the fusion characteristics of two shallow network layers in the fusion characteristics of the multiple layers to obtain a first convolution characteristic and a second convolution characteristic;

and performing addition operation on the first convolution characteristic and the second convolution characteristic to obtain an addition characteristic, and performing boundary guiding on the addition characteristic by adopting a plurality of convolution layers to obtain a boundary map.

5. The method for dividing a camouflage target image based on multi-level feature fusion according to claim 1, wherein the feature interaction is performed on the fusion features of adjacent network layers in the multi-level fusion features to obtain a plurality of interaction features, and the method comprises the following steps:

introducing a multi-scale channel attention mechanism into a feature interaction module, and adding fusion features of adjacent network layers in the multi-level fusion features to obtain a plurality of addition features;

inputting each added feature into the multi-scale channel attention mechanism to obtain a plurality of multi-scale channel features;

Obtaining a plurality of normalized features by adopting an activation function for each multi-scale channel feature, and obtaining a plurality of normalized difference features by subtracting each normalized feature;

performing feature enhancement on the plurality of normalized features and the plurality of normalized difference features to obtain a plurality of enhanced normalized features and a plurality of enhanced normalized difference features;

residual connection is carried out on each enhanced normalized feature and the corresponding fusion feature to obtain a plurality of first residual features, and each first residual feature is convolved to obtain a plurality of first convolution features;

residual connection is carried out on each enhanced normalized difference feature and the corresponding fusion feature to obtain a plurality of second residual features, and convolution is carried out on each second residual feature to obtain a plurality of second convolution features;

and adding each first convolution feature and the corresponding second convolution feature to obtain a plurality of added convolution features, and adopting a convolution layer to promote fusion of the added convolution features to obtain a plurality of interaction features.

6. The method for segmenting a camouflage target image based on multi-level feature fusion according to claim 1, wherein the performing boundary fusion on the boundary map and each of the plurality of interaction features to obtain a plurality of boundary fusion features includes:

Based on each interaction characteristic, learning a target overall characteristic by adopting a target attention head branch; wherein the target attention head branch is used for separating a target and a background on the whole based on the interaction characteristics;

based on the boundary map and each interaction characteristic, adopting a boundary attention head branch to learn boundary detail characteristics; the boundary attention head branches are used for capturing sparse local boundary information of the target based on the boundary map and the interaction characteristics;

splicing the output of each target attention head branch and the output of each boundary attention head branch corresponding to the target attention head branch to obtain a plurality of output splicing features, and adopting a convolution layer to promote feature fusion of the plurality of output splicing features to obtain a plurality of convolution fusion features;

and carrying out residual connection on each convolution fusion feature and each interaction feature corresponding to each convolution fusion feature to obtain a plurality of boundary fusion features.

7. The method for segmenting the camouflage target image based on multi-level feature fusion according to claim 1, wherein the segmenting the camouflage target image in the camouflage target image to be segmented corresponding to each boundary fusion feature based on the plurality of boundary fusion features comprises:

Inputting the boundary fusion features into a convolution layer with a Sigmoid activation function to generate a plurality of prediction graphs;

and dividing a camouflage target image in the camouflage target images to be divided based on each prediction graph.

8. The camouflage target image segmentation system based on the multi-level feature fusion is characterized by comprising:

the data acquisition unit is used for acquiring a camouflage target image to be segmented;

the feature extraction unit is used for extracting multi-level features of the camouflage target image to be segmented by adopting different network layers through a first branch network and a second branch network to obtain a first feature image of multiple levels output by the first branch network and a second feature image of multiple levels output by the second branch network;

the feature fusion unit is used for carrying out global feature enhancement on the first feature map of each level to obtain global features after multiple levels of enhancement; carrying out local feature enhancement on the second feature map of each level to obtain local features after multiple levels of enhancement; carrying out feature fusion on the enhanced local features of each level and the enhanced global features of the same level to obtain fusion features of multiple levels;

The boundary guiding unit is used for conducting boundary guiding on the fusion characteristics of the two shallow network layers in the fusion characteristics of the multiple layers to obtain a boundary diagram;

the feature interaction unit is used for carrying out feature interaction on the fusion features of the adjacent network layers in the fusion features of the multiple layers to obtain multiple interaction features;

the boundary fusion unit is used for carrying out boundary fusion on the boundary graph and each interaction feature in the interaction features respectively to obtain a plurality of boundary fusion features;

the image segmentation unit is used for segmenting a camouflage target image in the camouflage target images to be segmented, which correspond to each boundary fusion feature, based on the boundary fusion features.

9. An electronic device, comprising:

at least one memory;

at least one processor;

at least one computer program;

the at least one computer program is stored in the at least one memory, the at least one processor executing the at least one computer program to implement:

the camouflage target image segmentation method based on multi-level feature fusion according to any one of claims 1 to 7.

10. A storage medium that is a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program for causing a computer to execute:

A camouflage target image segmentation method based on multi-level feature fusion as claimed in any one of claims 1 to 7.