A multi-prior driven salient object detection algorithm

Abstract

The application provides a multi-prior-driven salient object detection algorithm, and relates to the technical field of image processing, and comprises the following steps: obtaining a to-be-detected target image, and preprocessing the to-be-detected target image to obtain a prior image with saliency prior knowledge; adopting prior visual attention modules with different network parameter amounts to process different prior images to obtain eight saliency feature maps with different scales; fusing the eight saliency feature maps with different scales to obtain four fusion feature representations; extracting salient object edge feature information according to the four fusion feature representations; supplementing the extracted salient object edge feature information to a full-resolution saliency mask after up-sampling, and completing target detection according to the saliency mask after supplementing information, so that the application can realize accurate segmentation of the edges of a main body in various complex environments, such as low foreground and background contrast, complex background and complex main body shape.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a multi-prior driven salient object detection algorithm. Background Art

[0002] In computer vision, a saliency mask is an image that shows the unique qualities of each pixel. The goal of a saliency mask is to simplify or change the image representation to one that is more meaningful and easier to analyze. For example, if a pixel has a high grayscale level or other unique color quality in a color image, the quality of that pixel will be more noticeable in the saliency mask. Saliency detection can be considered an example of image segmentation.

[0003] Salient object detection is primarily used for image foreground segmentation, object tracking, image understanding, and weakly supervised learning. It enables the rapid design of creative images, as well as the replacement of backgrounds within images or video frames, integrating foreground characters into different scenes. It also provides guidance for advanced visual tasks, driving the development of the intelligent industry. Traditional manual processing methods require specialized skills and are often labor-intensive, slow, and ineffective in obtaining manual priors. In recent years, with the development of deep learning algorithms, image semantic segmentation algorithms have matured, and segmentation algorithms based on salient objects have become widely used. However, these algorithms themselves also suffer from issues such as a lack of refined edge resolution and a bulky image feature extraction module.

[0004] In summary, how to overcome the above-mentioned defects is a problem that those skilled in the art urgently need to solve. Summary of the Invention

[0005] In response to the above-mentioned problems and needs, this solution proposes a multi-prior driven salient object detection algorithm. Specifically, it can solve the above-mentioned technical problems by adopting the following technical solutions.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a multi-prior driven salient object detection algorithm, comprising: obtaining an image of a target to be detected, and preprocessing the image of the target to be detected to obtain a prior image with saliency prior knowledge;

[0007] The prior visual attention modules with different network parameters are used to process different prior images to obtain saliency feature maps at eight different scales.

[0008] Performing feature fusion on the eight saliency feature maps of different scales to obtain four fused feature representations containing semantic information and structural detail information;

[0009] Extracting salient target edge feature information based on the four fused feature representations;

[0010] The extracted salient target edge feature information is added to the upsampled full-resolution saliency mask, and target detection is completed based on the saliency mask after supplementation.

[0011] Furthermore, the pre-processing process includes:

[0012] Acquire five prior images of the input target image to be detected by a machine learning method, wherein the five prior images include LC, FT, FG, RC, and Gradient;

[0013] Constructing a priori guidance network, and using it to select prior images with stronger saliency representation capabilities than the original input image, and performing loss calculation and backpropagation optimization on the final output saliency mask and label image. The priori guidance network consists of a ResNet50 network and a multi-prior attention network.

[0014] The three-channel RGB image is input into the ResNet50 network for processing to obtain first feature information, and the multi-prior attention network uses four spatial attention modules with grouped convolution to process the input three-channel prior feature map and output second feature information;

[0015] The first feature information and the second feature information are added and fused by the add function to output a priori feature map;

[0016] According to the prior feature map output by the prior guidance network, FG and Gradient are selected as the prior images with significant prior knowledge.

[0017] Furthermore, the a priori visual attention module with different network parameter amounts processes different a priori feature maps to obtain eight saliency feature maps of different scales, including four features obtained from Gradient by the a priori visual attention module, the four features being F G2 、F G4 、F G8 and F G16 , four features obtained from FG, the four features are F F2 、F F4 、F F8 and F F16 , and use the label mask to upsample to the original image resolution F G2 、F G4 、F G8 、F G16 、F F2 、F F4 、F F8 and F F16Perform loss calculation and stage-by-stage back-propagation optimization; the prior visual attention module is composed of four levels of pseudo-twin network blocks connected in series, that is, the input of the current-level pseudo-twin network block is the output of the previous-level pseudo-twin network block, and each pseudo-twin network block includes a channel configuration submodule, a spatial attention submodule and an inter-group integration submodule, and the three submodules are connected in series in sequence.

[0018] Furthermore, the channel configuration submodule increases the number of input feature channels of each level of pseudo twin network block to 64, 128, 256 and 512, thereby increasing and refining the significant feature information; the spatial attention submodule uses the grouped convolution in the PyTorch framework to perform grouping processing according to the number of channels, and the number of groups is consistent with the number of input feature map channels. There are three types of convolution kernel sizes k for the grouped convolution, k=1, k=5 and k=7, and the activation function uses Gaussian error linear unit GELU for feature activation; the inter-group integration submodule uses 2D convolution to process the features output by the spatial attention submodule, and converts the number of output feature channels to the number of feature channels required by the next level pseudo twin network block, and the output is F F2 、F G2 、F F4 、F G4 、F F8 、F G8 、F F16 and F G16 Eight characteristics.

[0019] Furthermore, the feature fusion of the eight saliency feature maps of different scales includes:

[0020] The eight input features are divided into four groups according to the resolution size through pixel-by-pixel addition operation, and the add function is used to fuse the two features with the same resolution in each group and perform batchnorm batch normalization and GELU linear activation;

[0021] Then, through the feature information supplementation operation, the detail information of the larger resolution feature is used to supplement the semantic information of the smaller resolution feature: F G2 Downsample to F G4 、F G8 and F G16 The resolution is added pixel by pixel to make the original image resolution F G4 Downsample to F G8 and F G16 , the original image resolution F G8 Downsample to F G16 , the resolution is added and fused pixel by pixel;

[0022] Use the label mask to upsample to the original image resolution F P2 、F P4 、F P8 and F P16 Perform loss calculation and staged backpropagation optimization.

[0023] Furthermore, the extracting of significant target edge feature information based on the four fused feature representations includes performing edge feature extraction on the four input features through an edge extraction operation, and then fusing the four features of different resolutions through an edge feature fusion operation to output a significant edge feature map F of full resolution. E ;

[0024] The edge extraction operation processes the input features using two sets of dilated convolutions with different convolution kernel sizes k (k=1, 3) and different dilation rates d (d=1, 2). An attention weight map is calculated based on the feature difference after the two sets of dilated convolutions. The original feature map is multiplied by the attention weight map to obtain a coarse-grained salient edge feature. The salient edge feature is then convolved, reshaped, and transposed to calculate a non-local enhancement weight map. The coarse-grained salient edge feature is multiplied by the non-local enhancement weight map to obtain a fine-grained salient edge feature map.

[0025] The edge feature fusion operation first upsamples the four obtained significant edge feature maps to the original image resolution, then uses the cat function in the PyTorch framework to splice them by channel, and finally uses a 2D convolution with a convolution kernel size of 3*3 to reduce the number of splicing feature channels to 1;

[0026] And use the edge label mask to E Perform loss calculation and staged backpropagation optimization.

[0027] Furthermore, the method of adding the extracted salient target edge feature information to the upsampled full-resolution saliency mask includes: firstly, the input F P2 、F P4 、F P8 and F P16 The feature map is upsampled to the original image resolution and fused into a saliency feature map F with channel number 1. P ; Then use the add function in PyTorch to add F P and F E The saliency mask is finally obtained by adding and fusion pixel by pixel.

[0028] As can be seen from the above technical solution, the beneficial effects of the present invention are: it can combine traditional machine learning image processing with deep learning, reducing the number of parameters in the deep learning network module and improving the model's computational speed. It also supplements the edges of the saliency mask with structural details, such as the edges, to accurately capture salient objects and further segment the salient subject from the background. In complex environments such as low foreground-background contrast, complex backgrounds, and complex subject shapes, it can achieve accurate subject edge segmentation with strong robustness.

[0029] In addition to the objects, features and advantages described above, the best embodiments for implementing the present invention will be described in more detail below with reference to the accompanying drawings so that the features and advantages of the present invention can be easily understood. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments of the present invention or the description of the prior art. The drawings are only used to show some embodiments of the present invention, rather than to limit all embodiments of the present invention thereto.

[0031] Figure 1 Schematic diagram of the specific steps of a multi-prior driven salient object detection algorithm in the present invention.

[0032] Figure 2 Schematic diagram of the network structure of the a priori guidance network according to an embodiment of the present invention.

[0033] Figure 3 Schematic diagram of the overall network structure of the salient object detection network according to an embodiment of the present invention.

[0034] Figure 4 Schematic diagram of the structure of the prior visual attention module of an embodiment of the present invention.

[0035] Figure 5 Schematic diagram of the feature fusion network structure of an embodiment of the present invention.

[0036] Figure 6 Schematic diagram of the edge feature extraction network structure according to an embodiment of the present invention.

[0037] Figure 7 Schematic diagram of the effect of salient object detection in different scenarios according to an embodiment of the present invention. DETAILED DESCRIPTION

[0038] In order to make the purpose, technical solution and advantages of the technical solution of the present invention clearer, the technical solution of the embodiment of the present invention will be clearly and completely described below in conjunction with the drawings of specific embodiments of the present invention. The same figure marks in the drawings represent the same components. It should be noted that the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the described embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0039] The present invention adopts a multi-prior driven salient object detection network, uses traditional machine learning methods to quickly obtain different priors of the input image, and guides the algorithm network to detect the foreground subject of the image. It can quickly and accurately capture the salient subject of the image and refine the edge details to improve wider applicability.

[0040] like Figures 1 to 2 As shown in Figure 2, a multi-prior driven salient object detection algorithm specifically includes the following steps:

[0041] Step 1: Acquire a target image to be detected and preprocess the target image to obtain a priori image with saliency prior knowledge;

[0042] Preprocessing involves converting the image to be processed into a feature map with significant prior knowledge using traditional digital image processing and traditional machine learning methods. Traditional digital image processing and traditional machine learning methods mainly use the OpenCV framework. The resolution and number of channels of the processed prior feature map are consistent with the original image.

[0043] Specifically, the pretreatment process includes:

[0044] Step 1.1: Obtain five prior images of the input target image to be detected by machine learning method, wherein the five prior images include LC, FT, FG, RC and Gradient;

[0045] Step 1.2: Construct a prior guidance network and use it to select a prior image with stronger saliency representation ability than the original input image. Perform loss calculation and backpropagation optimization on the final output saliency mask and label image. The prior guidance network consists of a ResNet50 network and a multi-prior attention network.

[0046] Step 1.3: The three-channel RGB image is input into the ResNet50 network for processing to obtain the first feature information. The multi-prior attention network uses four spatial attention modules with grouped convolution to process the input three-channel prior feature map and output the second feature information;

[0047] Step 1.4: Use the add function to add and fuse the first feature information and the second feature information to output a priori feature map;

[0048] Step 1.5: According to the prior feature map output by the prior guidance network, FG and Gradient are selected as prior images with significant prior knowledge.

[0049] In this embodiment, five prior images (including LC, FT, FG, RC and Gradient) of RGB images are obtained through traditional image processing or traditional machine learning methods encapsulated in OpenCV. A simple prior guidance network is designed to conduct ablation experiments to verify the strength of the saliency representation capabilities of the five prior images and the original input image, and select the prior image with stronger saliency representation capabilities than the original input image. Specifically, two effective saliency representation priors, Gradient and FG, are selected, and the saliency representation capability of FG is stronger than that of Gradient. The simple prior guidance network includes two modules: Backbone network and multi-prior attention network. Among them, the Backbone network uses the Residual Network subnetwork ResNet50 to process the input three-channel RGB image, and the multi-prior attention network uses four spatial attention modules with grouped convolution to process the input three-channel prior feature map. Finally, the full-resolution features obtained by the two modules are added and fused using the add function in the PyTorch framework and then output.

[0050] Step 2: Use the prior visual attention modules with different network parameters to process different prior images and obtain eight saliency feature maps of different scales.

[0051] Specifically including: four features obtained from Gradient through the prior visual attention module, the four features are F G2 、F G4 、F G8 and F G16 , four features obtained from FG, the four features are F F2 、F F4 、F F8 and F F16 , and use the label mask to upsample to the original image resolution F G2 、F G4 、F G8 、F G16 、F F2 、F F4 、F F8 and F F16Perform loss calculation and phased back propagation optimization; the prior visual attention module is composed of four levels of pseudo twin network blocks in series, that is, the input of the current level pseudo twin network block is the output of the previous level pseudo twin network block, each pseudo twin network block includes a channel configuration submodule, a spatial attention submodule and an inter-group integration submodule, and the three submodules are connected in series in sequence, wherein the channel configuration submodule increases the number of input feature channels of each level of pseudo twin network block to 64, 128, 256 and 512, thereby increasing and refining the significant feature information; the spatial attention The force submodule uses the group convolution in the PyTorch framework to perform group processing according to the number of channels. The number of groups is consistent with the number of channels in the input feature map. There are three types of convolution kernel sizes k for the group convolution, k=1, k=5 and k=7, and the activation function uses the Gaussian error linear unit GELU for feature activation; the inter-group integration submodule uses 2D convolution to process the features output by the spatial attention submodule, and converts the number of output feature channels to the number of feature channels required by the next-level pseudo-twin network block, which is convenient for subsequent forward series processing. The output is F F2 、F G2 、F F4 、F G4 、F F8 、F G8 、F F16 and F G16 Eight characteristics.

[0052] Using grouped convolution to process the features of each channel by channel is more conducive to focusing on the relevant features of the salient object. Therefore, after processing by the above module, redundant information in the prior feature map is further ignored, highlighting the relevant features of the salient object. Features with higher resolution contain more structural details, while features with lower resolution contain more overall semantic information. Therefore, the proposed module also has the function of classifying salient object feature information.

[0053] Step 3: Perform feature fusion on the eight saliency feature maps of different scales to obtain four fusion feature representations containing semantic information and structural detail information. Specifically, the eight input features are cascaded into four cumulative features F through pixel-by-pixel addition operation and feature information supplementation and enhancement operation. P2 、F P4 、F P8 and F P16 .

[0054] like Figure 5 The feature fusion network structure shown in FIG. 1 includes:

[0055] Step 3.1: Divide the eight input features into four groups according to the resolution size through pixel-by-pixel addition operation, use the add function in the PyTorch framework to fuse the two features with the same resolution in each group, and perform batchnorm batch normalization and GELU linear activation;

[0056] Step 3.2: Then, through the feature information supplementation operation, the detail information of the larger resolution feature is used to supplement the semantic information of the smaller resolution feature: F G2 Downsample to F G4 、F G8 and F G16 The resolution is added pixel by pixel to make the original image resolution F G4 Downsample to F G8 and F G16 The resolution is added pixel by pixel to make the original image resolution F G8 Downsample to F G16 The pixel-by-pixel addition and fusion operation can make all the accumulated features have a certain degree of significant subject detail information;

[0057] Step 3.3: Use the label mask to upsample to the original image resolution F P2 、F P4 、F P8 and F P16 Perform loss calculation and staged backpropagation optimization.

[0058] The eight feature maps of different scales are gradually integrated, so the salient object information is gradually accumulated. First, four sets of features with the same resolution are added together using pixel-by-pixel addition to obtain four fused features. Then, the larger resolution features are downsampled to enhance the smaller resolution features, so that the smaller resolution features also contain detailed information. Therefore, the accumulated feature information not only contains the semantic information of the entire object, but also contains structural details that are helpful for the edge prior capture module.

[0059] Step 4: Extract the edge feature information of the salient target based on the four fused feature representations.

[0060] like Figure 4 The edge feature extraction network structure shown in the figure can use convolution groups with different convolution kernel sizes and different void ratios to process accumulated features of different scales, gradually refine the edge structured information, and use it to supplement the full-resolution saliency mask after the final upsampling.

[0061] Specifically, the extracting of significant target edge feature information based on the four fused feature representations includes performing edge feature extraction on the four input features through an edge extraction operation, and then fusing the four features of different resolutions using an edge feature fusion operation to output a significant edge feature map F of full resolution. E ;

[0062] Step 4.1: The edge extraction operation uses two sets of dilated convolution groups with different convolution kernel sizes k (k = 1, 3) and different dilation rates d (d = 1, 2) to process the input features. The attention weight map is calculated based on the feature difference after the two sets of dilated convolutions. The original feature map is multiplied by the attention weight map to obtain a coarse-grained salient edge feature. The salient edge feature is then subjected to convolution, reshape, and transposition operations so that each pixel can fully consider the relationship with all other pixels to calculate the non-local enhancement weight map. The coarse-grained salient edge feature is multiplied by the non-local enhancement weight map to obtain a fine-grained salient edge feature map.

[0063] Step 4.2: The edge feature fusion operation first upsamples the four obtained significant edge feature maps to the original image resolution, then uses the cat function in the PyTorch framework to splice them by channel, and finally uses a 2D convolution with a convolution kernel size of 3*3 to reduce the number of splicing feature channels to 1;

[0064] Step 4.3: And use the edge label mask to F E Perform loss calculation and staged backpropagation optimization.

[0065] Step 5: The extracted edge feature information of the salient object is added to the upsampled full-resolution saliency mask, and the object detection is completed based on the supplemented saliency mask. Therefore, the final saliency mask contains both the semantic information of the salient object and more structural details.

[0066] Step 5 specifically includes:

[0067] Step 5.1: First, input F P2 、F P4 、F P8 and F P16 The feature map is upsampled to the original image resolution and fused into a saliency feature map F with channel number 1. P ;

[0068] Step 5.2: Use the add function in PyTorch to add F P and F containing edge detail information EThe final saliency mask is obtained by pixel-by-pixel addition and fusion. The saliency mask obtained by this operation takes into account both the semantic information and detail information of the salient object.

[0069] like Figure 3 As shown, the present invention uses mature traditional image processing methods or traditional machine learning methods to obtain a priori feature maps of the image to be detected, and uses the prior features to design a salient object detection network with a small number of model parameters and fast inference speed, guiding the network to combine the semantic information and structured detail information of the salient object to obtain a more refined saliency mask. Specifically, it includes using traditional digital image processing methods or traditional machine learning methods encapsulated by OpenCV to obtain five prior feature maps of the input RGB image, and conducting ablation experiments to select two feature maps with strong saliency representation capabilities for the input of the salient target detection network; using prior visual attention modules with different network parameter values ​​to process different prior feature maps, remove redundant information in the prior feature maps, and highlight the relevant features of the salient target body; then gradually integrating the eight feature maps of different scales obtained from the prior visual attention module, so that the accumulated feature information not only contains the semantic information of the target as a whole, but also contains structural detail information that is helpful for the edge prior capture part; the edge prior capture part is used to extract the salient target edge features from the acquired features, which are used to supplement the full-resolution saliency mask after the final upsampling; finally, the obtained features are fused with the edge prior capture part using a pixel-by-pixel addition output operation to output the final full-resolution saliency mask.

[0070] Depend on Figure 7 The renderings of salient target detection in different scenarios show that the present invention is capable of performing high-precision salient target detection on targets in complex scenarios such as multiple targets, target occlusion, and tiny targets, and obtaining high-quality foreground subject segmentation result maps. It can achieve accurate segmentation of subject edges in various complex environments such as low contrast between foreground and background, complex background, and complex subject shapes.

[0071] It should be noted that the embodiments described in the present invention are only preferred ways to implement the present invention, and any obvious modifications that belong to the overall concept of the present invention should fall within the scope of protection of the present invention.

Claims

1. A multi-prior driven salient object detection algorithm, characterized by: The following steps are involved: Acquire a target image to be detected, and preprocess the target image to be detected to obtain a priori image with saliency prior knowledge; The prior visual attention modules with different network parameters are used to process different prior images to obtain saliency feature maps at eight different scales. Performing feature fusion on the eight saliency feature maps of different scales to obtain four fused feature representations containing semantic information and structural detail information; Extracting salient target edge feature information based on the four fused feature representations; The extracted salient target edge feature information is added to the upsampled full-resolution saliency mask, and the target detection is completed based on the saliency mask after the supplementary information; The pretreatment process includes: Acquire five prior images of the input target image to be detected by a machine learning method, wherein the five prior images include LC, FT, FG, RC, and Gradient; Constructing a priori guidance network, and using it to select prior images with stronger saliency representation capabilities than the original input image, and performing loss calculation and backpropagation optimization on the final output saliency mask and label image. The priori guidance network consists of a ResNet50 network and a multi-prior attention network. The three-channel RGB image is input into the ResNet50 network for processing to obtain first feature information, and the multi-prior attention network uses four spatial attention modules with grouped convolution to process the input three-channel prior feature map and output second feature information; The first feature information and the second feature information are added and fused by the add function to output a priori feature map; According to the prior feature map output by the prior guidance network, FG and Gradient are selected as prior images with significant prior knowledge; The a priori visual attention module with different network parameter amounts is used to process different a priori feature maps to obtain eight saliency feature maps of different scales, including four features obtained from Gradient by the a priori visual attention module, the four features being 、 、 and , four features obtained from FG, the four features are 、 、 and , and use the label mask to upsample to the original image resolution 、 、 、 、 、 、 and Perform loss calculation and stage-by-stage back-propagation optimization; the prior visual attention module is composed of four levels of pseudo-twin network blocks connected in series, that is, the input of the current-level pseudo-twin network block is the output of the previous-level pseudo-twin network block, and each pseudo-twin network block includes a channel configuration submodule, a spatial attention submodule and an inter-group integration submodule, and the three submodules are connected in series in sequence.

2. The multi-prior driven salient object detection algorithm according to claim 1, characterized in that: The channel configuration submodule increases the number of input feature channels of each level of pseudo-twin network block to 64, 128, 256 and 512, thereby increasing and refining the significant feature information; The spatial attention submodule uses the grouped convolution in the PyTorch framework to perform grouping processing according to the number of channels. The number of groups is consistent with the number of channels in the input feature map. There are three types of convolution kernel sizes k for the grouped convolution, k=1, k=5 and k=7. The activation function uses the Gaussian error linear unit GELU for feature activation; the inter-group integration submodule uses 2D convolution to process the features output by the spatial attention submodule, and converts the number of feature channels output to the number of feature channels required by the next-level pseudo-twin network block. The output is 、 、 、 、 、 、 and Eight characteristics.

3. The multi-prior driven salient object detection algorithm according to claim 2, characterized in that: The feature fusion of the eight saliency feature maps at different scales includes: The eight input features are divided into four groups according to the resolution size through pixel-by-pixel addition operation, and the add function is used to fuse the two features with the same resolution in each group and perform batchnorm batch normalization and GELU linear activation; Then, through the feature information supplementation operation, the detail information of the larger resolution feature is used to supplement the semantic information of the smaller resolution feature: of Downsample to 、 and The resolution is added pixel by pixel to make the original image resolution of Downsample to and , the original image resolution of Downsample to , the resolution is added and fused pixel by pixel; Use label mask to upsample to original image resolution 、 、 and Perform loss calculation and staged backpropagation optimization.

4. The multi-prior driven salient object detection algorithm according to claim 3, wherein: The extracting of significant target edge feature information based on the four fused feature representations includes performing edge feature extraction on the four input features through an edge extraction operation, and then fusing the four features of different resolutions through an edge feature fusion operation to output a significant edge feature map of full resolution. ; The edge extraction operation processes the input features using two sets of dilated convolutions with different convolution kernel sizes k (k=1, 3) and different dilation rates d (d=1, 2). An attention weight map is calculated based on the feature difference after the two sets of dilated convolutions. The original feature map is multiplied by the attention weight map to obtain a coarse-grained salient edge feature. The salient edge feature is then convolved, reshaped, and transposed to calculate a non-local enhancement weight map. The coarse-grained salient edge feature is multiplied by the non-local enhancement weight map to obtain a fine-grained salient edge feature map. The edge feature fusion operation first upsamples the four obtained significant edge feature maps to the original image resolution, then uses the cat function in the PyTorch framework to splice them by channel, and finally uses a 2D convolution with a convolution kernel size of 3*3 to reduce the number of splicing feature channels to 1; And use the edge label mask to Perform loss calculation and staged backpropagation optimization.

5. The multi-prior driven salient object detection algorithm according to claim 4, characterized in that: The method of adding the extracted salient target edge feature information to the upsampled full-resolution saliency mask includes: firstly, 、 、 and The feature map is upsampled to the original image resolution and fused into a saliency feature map with channel number 1. ; Then use the add function in PyTorch to and The saliency mask is finally obtained by adding and fusion pixel by pixel.

Images