Method for detecting salient objects in optical remote sensing images based on edge-guided enhancement network

Abstract

The application discloses an optical remote sensing image salient object detection method based on an edge guidance enhanced network, which comprises the steps of image preprocessing, generating a predicted edge map tensor E, establishing an EGE network, and generating a saliency map using the EGE network. The application makes full use of edge information, accurately identifies the object boundary using a lightweight network, and generates a predicted edge map. The processed backbone features are gradually fused between high-level and low-level, the backbone features are concentrated on the target area, and interference sources are excluded. These fused features are integrated into deeper features, missing target information in the background is explored through a reverse attention mechanism, and missing target details in the background are better extracted.

Description

TECHNICAL FIELD

[0001] The present application relates to an optical remote sensing image salient object detection method, in particular to an optical remote sensing image salient object detection method based on edge guided enhancement network, and belongs to the technical field of computer vision. BACKGROUND

[0002] Salient object detection (SOD) is a fundamental task in computer vision, which is used to identify prominent elements in an image and improve the performance of various applications such as image segmentation, object detection, etc. Existing SOD methods focus on natural scenes and RGB-T, RGB-D images. With the development of earth observation technology, a large number of remote sensing images are generated, which are the product of information obtained by various sensors and the information carrier of remote sensing detection targets, and can distinguish various details such as water and vegetation. Optical remote sensing images (ORSIs) cover a wide area of high altitude areas, and the imaging conditions are complex, containing different types of objects and backgrounds.

[0003] Traditional methods usually rely on hand-crafted features or prior knowledge, such as sparse representation, color, superpixel, etc. Hand-crafted features show limited robustness to the diversity of objects, especially in complex and cluttered scenes. With the development of deep learning, CNN-based methods have become popular in ORSI-SOD. Early deep SOD models usually use a multi-layer perceptron (MLP) classifier to predict saliency scores from deep features. Later, CNN-based networks emerged, including a nested network that fuses multi-resolution features, a parallel downlink fusion network that is densely connected, and a dense attention flow network that generates shallow cues to guide the generation of high-level feature maps. To improve edge detection, some methods utilize multi-scale deep features with explicit and implicit edge information. Others utilize edge-guided recurrent networks with a parallel decoder structure. Although these methods recognize the importance of edge information, they cannot effectively combine edge information with backbone features. In addition, they usually rely on the same structure to extract object boundaries and structural information. However, object structure and edge extraction are different tasks, each of which is crucial to the effectiveness of SOD. Therefore, processing object structure and edge extraction separately will help improve the performance of optical remote sensing image salient object detection. SUMMARY

[0004] The present application aims to provide an optical remote sensing image salient object detection method based on edge guided enhancement network.

[0005] To solve the above technical problems, the technical scheme adopted by the present application is:

[0006] An optical remote sensing image salient object detection method based on edge guided enhancement network, comprising the following steps:

[0007] Step 1: image preprocessing: resize the input image to a preset size tensor X;

[0008] Step 2: input the tensor X into the edge generation network EGEUNet to generate a predicted edge map tensor E;

[0009] Step 3: establish the EGE network: the EGE network includes a cascaded encoder, an intermediate layer and a decoder, the encoder includes one or more feature encoders, a pre-positioning module, the intermediate layer includes adjacent context enhancement modules with the same number as the feature encoders, and the decoding layer includes cross-feature fusion modules and reverse refinement modules; the number of cross-feature fusion modules and reverse refinement modules is the same, and both are less than the number of feature encoders by 1;

[0010] The pre-positioning module PPM is located after the designated feature encoder and is sequentially connected with each feature encoder. The output end of each feature encoder is connected with the corresponding adjacent context enhancement module ACEM of the adjacent feature encoder. The output end of each adjacent context enhancement module ACEM is connected with the corresponding input end of the corresponding cross-feature fusion module CFF. The output end of the last-stage adjacent context enhancement module ACEM is connected with the output end of the Conv module to output the last-stage predicted map. The reverse refinement module RRM receives the next-stage predicted map and the output of the cross-feature fusion module CFF of the current stage to generate the predicted map of the current stage. The first-stage predicted map is output as a saliency map.

[0011] Step 4: generate a saliency map using the EGE network: input the tensor X and the predicted edge map tensor E into the EGE network, and after processing by the encoder, the intermediate layer and the decoder, output the predicted Figure 1 predicted Figure 5 predicted Figure 1 saliency map.

[0012] Further, the information processing process of the pre-positioning module PPM is as follows:

[0013] f’3=Max(Norm(Conv g (E)))*f3+f3

[0014] where Conv g (·) is a Gaussian convolution kernel, Norm(·) is a normalization process, Max(·) is a maximum value calculation. f3 is the feature tensor output by the feature encoder, and f’3 is the feature tensor output after processing by the pre-positioning module PPM.

[0015] Further, each adjacent context enhancement module ACEM in the middle stage has the same structure, including a local branch, a first and a second adjacent branch, and a fusion module; the local branch is composed of a cascaded channel attention module CA and a receptive field module RFB; the first adjacent branch is composed of a cascaded downsampling module D and a spatial attention module SA1; the second adjacent branch is composed of a cascaded upsampling module U and a second spatial attention module SA2; the fusion module is composed of a first and a second multiplier and an adder; the output of the receptive field module RFB is connected to the input of the first and the second multiplier and the adder, and the input of the adder is connected to the input of the local branch and the first and the second multiplier;

[0016] The structure of the adjacent context enhancement module in the first and the last stage is different from that in the middle stage, in that the adjacent context enhancement module in the first stage includes a local branch and a first adjacent branch, and the local branch is composed of a RFB module and does not include a cascaded channel attention module CA; the adjacent context enhancement module in the last stage includes a local branch and a second adjacent branch, and the local branch is composed of a RFB module and does not include a cascaded channel attention module CA.

[0017] Further, the cross-feature fusion module CFF includes convolution kernels CBR1-CBR8, the output data of the convolution kernel CBR1 is multiplied by the convolution kernel CBR8, one of the results is input into the splicing unit after being processed by the convolution kernel CBR7, and the other is input into the splicing unit after being processed by the convolution kernels CBR1-CBR3 in sequence, the output of the convolution kernel CBR3 is input into the splicing unit after being processed by the convolution kernel CBR8, the output of the splicing unit is added to the output after being processed by the convolution kernel CBR6 and multiplied by the predicted edge map tensor E, and the output is the fused feature The convolution kernels CBR1-CBR6 have the same structure, each of which is composed of a 3x3 convolution layer, a batch normalization layer and a ReLU activation layer, the dilation rate of the convolution kernel CBR1 is 1, the convolution kernel CBR7 is different from the convolution kernel CBR1 in that the dilation rate is 2, and the convolution kernel CBR8 is different from the convolution kernel CBR1 in that the dilation rate is 3.

[0018] Further, the reverse refinement module RRM is composed of three branches, and the output prediction map p t is obtained by element-level summation of the three branches;

[0019]

[0020] wherein, is the feature map output by the ACEM, Conv(·) is a convolution layer with a kernel size of 1x1, is the output of the second branch, is the output of the third branch, and ⊕ represents element-level summation.

[0021] The first branch is input to the next level prediction map;

[0022] The second branch is expressed as:

[0023]

[0024] The third branch is expressed as:

[0025]

[0026] where CBRG represents three CBR convolutional layers, the first two convolutional kernel sizes are 3x3, the channel compression number is 1, and the last kernel is 1x1, p i represents the prediction map generated by the current level, and σ(·) represents a sigmoid activation function, represents an output tensor.

[0027] Further, the encoder in the EGE network comprises feature encoders RES1-RES5, a pre-positioning module PPM, the intermediate layer comprises adjacent context enhancement modules ACEM1-ACEM5, the decoding layer comprises cross-feature fusion modules CFF1-CFF4 and reverse refinement modules RRM1-RRM4; the feature encoders RES1-RES3, the pre-positioning module PPM, the feature encoders RES4-RES5 are sequentially cascaded, the input ends of the adjacent context enhancement module ACEM1 are respectively connected with the corresponding output ends of the feature encoders RES1 and RES2, the output end thereof is connected with the corresponding input end of the cross-feature fusion module CFF1, the input ends of the adjacent context enhancement module ACEM2 are respectively connected with the corresponding output ends of the feature encoders RES1, RES2 and the pre-positioning module PPM, the output end thereof is connected with the corresponding input end of the cross-feature fusion module CFF2, the input ends of the adjacent context enhancement module ACEM3 are respectively connected with the corresponding output ends of the feature encoders RES2, the pre-positioning module PPM and the feature encoders RES4, the output end thereof is connected with the corresponding input end of the cross-feature fusion module CFF3, the input ends of the adjacent context enhancement module ACEM4 are respectively connected with the corresponding output ends of the pre-positioning module PPM, the feature encoders RES4 and RES5, the output end thereof is connected with the corresponding input end of the cross-feature fusion module CFF4, the input ends of the adjacent context enhancement module ACEM5 are respectively connected with the corresponding output ends of the feature encoders RES4 and RES5, the output end thereof is connected with the corresponding input end of the cross-feature fusion module CFF4, the corresponding output end of the cross-feature fusion module CFF4 is respectively connected with the corresponding input end of the cross-feature fusion module CFF3 and the corresponding input end of the reverse refinement module RRM4, the corresponding output end of the cross-feature fusion module CFF3 is respectively connected with the corresponding input end of the cross-feature fusion module CFF2 and the corresponding input end of the reverse refinement module RRM3, the corresponding output end of the cross-feature fusion module CFF2 is respectively connected with the corresponding input end of the cross-feature fusion module CFF1 and the corresponding input end of the reverse refinement module RRM2, the corresponding output end of the cross-feature fusion module CFF1 is connected with the corresponding input end of the reverse refinement module RRM1, the input end of the Conv module is connected with the output end of the adjacent context enhancement module ACEM2, the output end thereof is connected with the corresponding input end of the reverse refinement module RRM4, the output end of the reverse refinement module RRM4 is connected with the corresponding input end of the reverse refinement module RRM3, the output end of the reverse refinement module RRM3 is connected with the corresponding input end of the reverse refinement module RRM2, the output end of the reverse refinement module RRM2 is connected with the corresponding input end of the reverse refinement module RRM1; the reverse refinement modules RRM1-RRM5 output predictions Figure 1 - predictions Figure 5 , predictions Figure 1 for a saliency map.

[0028] By adopting the technical scheme, the present application achieves the following technical effects:

[0029] 1. The present application makes full use of edge information, uses lightweight network to accurately identify object boundary and generate predicted edge map.

[0030] 2. The present application designs a positioning module PPM guided by predicted edge map in the encoder, concentrates main features on target area and excludes interference source.

[0031] 3. The adjacent context enhancement module ACEM proposed in the present application can capture comprehensive context and multi-scale information and transmit them to the decoder.

[0032] 4. The present application uses the cross feature fusion module CFF in the decoder to fuse high-level and low-level features, and through the reverse refinement module RRM, the significance map is gradually refined through reverse attention, and the missing target details in the background are better extracted. BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1 is a structural diagram of embodiment 1 of the present application.

[0034] Figure 2 is a structural diagram of the adjacent context enhancement module ACEM of embodiment 1 of the present application.

[0035] Figure 3 is a structural diagram of the cross feature fusion module CFF of embodiment 1 of the present application.

[0036] Figure 4 is a structural diagram of the reverse refinement module RRM of embodiment 1 of the present application.

[0037] Figure 5 is an input image of embodiment 1 of the present application.

[0038] Figure 6 is a significance map detected by embodiment 1 of the present application. DETAILED DESCRIPTION

[0039] The following examples are used to illustrate the present application.

[0040] Embodiment 1

[0041] Referring to Figure 1 , an edge-guided enhancement network method based on significant target detection of optical remote sensing image, comprising the following steps:

[0042] Step 1: Data preprocessing: resize the input image to a preset size tensor X, input the tensor X into the edge generation network EGEUNet to generate a predicted edge map tensor E; the edge generation network adopts the technical solution in the literature "Ruan, J., Xie, M., Gao, J., Liu, T., Fu, Y. (2023). EGE-UNet: An Efficient Group Enhanced UNet for Skin Lesion Segmentation. In: Greenspan, H., et al. Medical Image Computing and Computer Assisted Intervention - MICCAI 2023. MICCAI 2023. Lecture Notes in Computer Science, vol 14223. Springer, Cham."

[0043] Step 2: input the tensor X into the edge generation network EGEUNet to generate a predicted edge map tensor E;

[0044] Step 3: Establishing the EGE network: the EGE network includes cascaded encoders, intermediate layers and decoders, the encoders include feature encoders RES1-RES5, a pre-positioning module PPM, the intermediate layers include adjacent context enhancement modules ACEM1-ACEM5, and the decoding layers include cross-feature fusion modules CFF1-CFF4 and reverse refinement modules RRM1-RRM4; the feature encoders RES1-RES3, the pre-positioning module PPM, the feature encoders RES4-RES5 are sequentially cascaded, the input ends of the adjacent context enhancement module ACEM1 are respectively connected to the corresponding output ends of the feature encoders RES1 and RES2, the output end thereof is connected to the corresponding input end of the cross-feature fusion module CFF1, the input ends of the adjacent context enhancement module ACEM2 are respectively connected to the corresponding output ends of the feature encoders RES1, RES2 and the pre-positioning module PPM, the output end thereof is connected to the corresponding input end of the cross-feature fusion module CFF2, the input ends of the adjacent context enhancement module ACEM3 are respectively connected to the corresponding output ends of the feature encoders RES2, the pre-positioning module PPM and the feature encoders RES4, the output end thereof is connected to the corresponding input end of the cross-feature fusion module CFF3, the input ends of the adjacent context enhancement module ACEM4 are respectively connected to the corresponding output ends of the pre-positioning module PPM, the feature encoders RES4 and RES5, the output end thereof is connected to the corresponding input end of the cross-feature fusion module CFF4, the input ends of the adjacent context enhancement module ACEM5 are respectively connected to the corresponding output ends of the feature encoders RES4 and RES5, the output end thereof is connected to the corresponding input end of the cross-feature fusion module CFF4, the corresponding output end of the cross-feature fusion module CFF4 is respectively connected to the corresponding input end of the cross-feature fusion module CFF3 and the corresponding input end of the reverse refinement module RRM4, the corresponding output end of the cross-feature fusion module CFF3 is respectively connected to the corresponding input end of the cross-feature fusion module CFF2 and the corresponding input end of the reverse refinement module RRM3, the corresponding output end of the cross-feature fusion module CFF2 is respectively connected to the corresponding input end of the cross-feature fusion module CFF1 and the corresponding input end of the reverse refinement module RRM2, the corresponding output end of the cross-feature fusion module CFF1 is connected to the corresponding input end of the reverse refinement module RRM1, the input end of the Conv module is connected to the output end of the adjacent context enhancement module ACEM2, the output end thereof is connected to the corresponding input end of the reverse refinement module RRM4, the output end of the reverse refinement module RRM4 is connected to the corresponding input end of the reverse refinement module RRM3, the output end of the reverse refinement module RRM3 is connected to the corresponding input end of the reverse refinement module RRM2, the output end of the reverse refinement module RRM2 is connected to the corresponding input end of the reverse refinement module RRM1; the reverse refinement modules RRM1-RRM5 output predictions Figure 1 - predictions Figure 5 .

[0045] The feature encoders RES1-RES5 output first to fifth features respectively, and the pre-positioning module PPM refines the third feature f3 using the predicted edge map P as a notice prompt. The pre-positioning module PPM aims to highlight the position of a significant target object and suppress irrelevant information, and the information processing process is represented as:

[0046] f'3 = Max(Norm(Conv g (P)))*f3 + f3

[0047] where Conv g (·) is a Gaussian convolution kernel, Norm(·) is a normalization process, and Max(·) is a maximum value calculation. f'3 is the output feature tensor processed by the pre-positioning module PPM.

[0048] The adjacent context enhancement modules ACEM2-ACEM4 are of the same structure, including a local branch, first and second adjacent branches, and a fusion module; the local branch is composed of a cascaded channel attention module CA and a receptive field block module RFB; the first adjacent branch is composed of a cascaded downsampling module D and a spatial attention module SA1; the second adjacent branch is composed of a cascaded upsampling module U and a second spatial attention module SA2; and the fusion module is composed of first and second multipliers and an adder. The output end of the receptive field block module RFB is connected to the input ends of the first and second multipliers and the adder, and the input end of the adder is connected to the input end of the local branch and the input ends of the first and second multipliers.

[0049] The adjacent context enhancement modules ACEM1 and ACEM5 are of the same structure, and differ from the adjacent context enhancement modules ACEM2-ACEM4 in that the local branch is composed of an RFB module and does not include a cascaded channel attention module CA. The adjacent context enhancement module ACEM1 includes a first adjacent branch and a local branch, and the adjacent context enhancement module ACEM5 includes a second adjacent branch and a local branch.

[0050] The receptive field block module RFB can perceive multi-scale objects from different receptive fields while reducing parameters and increasing non-linear depth, and the implementation is described in the document "Receptive Field Block Net for Accurate and Fast Object Detection. European Conference on Computer Vision, 2017".

[0051] The convolution module includes a batch normalization layer and a ReLU activation layer, which are used to reduce channels.

[0052] Since the features at this level are different in scale from two adjacent features and have different locations of interest in space, adjacent branches process adjacent features, and after down-sampling and up-sampling are performed on the left and right branches respectively, two adjacent levels of attention maps are obtained through a spatial attention module SA, and then the two attention maps are respectively multiplied with local branches at an element level to cross-align and highlight the region.

[0053] In the cross-feature fusion module CCF, the output data of the convolution kernel CBR1 and the convolution kernel CBR8 are multiplied, one of which is input into the splicing unit after being processed by the convolution kernel CBR7, and the other is input into the splicing unit after being processed by the convolution kernels CBR1-CBR3 in turn, the output of the convolution kernel CBR3 is input into the splicing unit after being processed by the convolution kernel CBR8, and the output of the splicing unit is added to the output after the output of the prediction edge map tensor E is multiplied by the output after being processed by the convolution kernel CBR6, to output the fused features The convolution kernels CBR1-CBR6 have the same structure, which consists of a 3x3 convolution layer, a batch normalization layer and a ReLU activation layer, the dilation rate of the convolution kernel CBR1 is 1, and the difference between the convolution kernel CBR7 and the convolution kernel CBR1 is that the dilation rate is 2, and the difference between the convolution kernel CBR8 and the convolution kernel CBR1 is that the dilation rate is 3.

[0054] The reverse refinement module RRM is applied after each CFF module. The reverse refinement module RRM generates an inverse attention map from the high-level feature prediction map, and combines it with the current feature map to explore the missing target clues in the background. The RRM consists of three branches. The obtained prediction map p t is obtained by element-level summation of the three branches.

[0055] The first branch is the prediction map of the previous level p t+1 The prediction map of p5 is directly obtained from . It is denoted as

[0056]

[0057] wherein, is the feature map output by ACEM, Conv(·) is a convolution layer with a kernel size of 1x1, is the output of the second branch, is the output of the third branch, and represents element-level summation.

[0058] The second branch of the RRM first connects the prediction map of the previous layer with the current feature. Then the combined feature map is processed by a cascaded convolution layer to integrate deeper features and gradually correct and refine the feature map. This process can be represented as:

[0059]

[0060] The third branch realizes the reverse attention by using the prediction map of the previous layer. An inverse attention map is generated and then multiplied with the current feature map. This combined map is then processed by a cascade of convolutional layers, allowing for a deeper exploration of the object information hidden in the background. This process can be represented as follows:

[0061]

[0062] where CBRG represents three CBR convolutional layers, the first two with a kernel size of 3x3. The number of compressed channels is 1, and the last kernel is 1x1. i represents the prediction map generated at each stage, and σ(·) represents the sigmoid activation function, represents the output tensor in the CFF module.

[0063] Step 4: Generate a saliency map using the EGE network: The tensor X and the predicted edge map tensor E are input into the EGE network, processed through the encoder, intermediate layer, and decoder, and output the predicted Figure 1 prediction Figure 5 prediction Figure 1 saliency map.

[0064] In this embodiment, the input image size is adjusted to a preset size tensor X, and the tensor size in this embodiment is 3x224x224; After processing by adjacent context enhancement modules ACEM1-ACEM5, the number of feature channels is compressed to 64, and the first to fifth feature tensors are obtained, with sizes of 64x224x224, 64x112x112, 64x56x56, 64x28x28, and 64x14x14, respectively. The sizes of the first to fifth prediction maps p1-p5 are 1x224x224, 1x112x112, 1x56x56, 1x28x28, and 1x14x14, respectively; The final result map uses the first layer prediction map as the saliency map output.

[0065] It should be noted that the technical scheme of the present application has been applied in a small range, and the research results show that the user satisfaction is high. Now it has begun to prepare for technical transformation application, and at the same time, intellectual property risk early warning research has been carried out.

Claims

1.A method for detecting salient objects in optical remote sensing images based on edge-guided enhancement network, characterized in that: The method comprises the following steps: Step 1: image preprocessing: adjusting the size of the input image to a preset size tensor X; Step 2: inputting the tensor X into an edge generation network EGEUNet to generate a predicted edge map tensor E; Step 3: establishing an EGE network: the EGE network comprises cascaded encoders, intermediate layers and decoders, the encoders comprise one or more feature encoders and a pre-positioning module, the intermediate layers comprise adjacent context enhancement modules which are the same in number as the feature encoders, and the decoding layers comprise cross-feature fusion modules and reverse refinement modules; The number of cross-feature fusion modules and reverse refinement modules is the same, and each is less than the number of feature encoders by 1; The pre-positioning module PPM is located behind the designated feature encoder and is sequentially cascaded with each feature encoder, the output ends of each feature encoder and the adjacent feature encoder are connected to the corresponding adjacent context enhancement module ACEM, the output end of each adjacent context enhancement module ACEM is connected to the corresponding input end of the corresponding cross-feature fusion module CFF, the output end of the last-stage adjacent context enhancement module ACEM is connected to the output end of the Conv module to output the last-stage predicted map; the reverse refinement module RRM receives the next-stage predicted map and the output of the cross-feature fusion module CFF of the current stage to generate the predicted map of the current stage; and the first-stage predicted map is output as a saliency map. The encoder in the EGE network comprises feature encoders RES1-RES5, a pre-positioning module PPM, the intermediate layer comprises adjacent context enhancement modules ACEM1-ACEM5, and the decoding layer comprises cross-feature fusion modules CFF1-CFF4 and reverse refinement modules RRM1-RRM4; the feature encoders RES1-RES3, the pre-positioning module PPM, the feature encoders RES4-RES5 are sequentially cascaded, the input ends of the adjacent context enhancement module ACEM1 are respectively connected to the corresponding output ends of the feature encoders RES1 and RES2, the output end thereof is connected to the corresponding input end of the cross-feature fusion module CFF1, the input ends of the adjacent context enhancement module ACEM2 are respectively connected to the corresponding output ends of the feature encoders RES1, RES2 and the pre-positioning module PPM, the output end thereof is connected to the corresponding input end of the cross-feature fusion module CFF2, the input ends of the adjacent context enhancement module ACEM3 are respectively connected to the corresponding output ends of the feature encoders RES2, the pre-positioning module PPM and the feature encoders RES4, the output end thereof is connected to the corresponding input end of the cross-feature fusion module CFF3, the input ends of the adjacent context enhancement module ACEM4 are respectively connected to the corresponding output ends of the pre-positioning module PPM, the feature encoders RES4 and RES5, the output end thereof is connected to the corresponding input end of the cross-feature fusion module CFF4, the input ends of the adjacent context enhancement module ACEM5 are respectively connected to the corresponding output ends of the feature encoders RES4 and RES5, the output end thereof is connected to the corresponding input end of the cross-feature fusion module CFF4, the corresponding output end of the cross-feature fusion module CFF4 is respectively connected to the corresponding input end of the cross-feature fusion module CFF3 and the corresponding input end of the reverse refinement module RRM4, the corresponding output end of the cross-feature fusion module CFF3 is respectively connected to the corresponding input end of the cross-feature fusion module CFF2 and the corresponding input end of the reverse refinement module RRM3, the corresponding output end of the cross-feature fusion module CFF2 is respectively connected to the corresponding input end of the cross-feature fusion module CFF1 and the corresponding input end of the reverse refinement module RRM2, the corresponding output end of the cross-feature fusion module CFF1 is connected to the corresponding input end of the reverse refinement module RRM1, the input end of the Conv module is connected to the output end of the adjacent context enhancement module ACEM5, the output end thereof is connected to the corresponding input end of the reverse refinement module RRM4, the output end of the reverse refinement module RRM4 is connected to the corresponding input end of the reverse refinement module RRM3, the output end of the reverse refinement module RRM3 is connected to the corresponding input end of the reverse refinement module RRM2, the output end of the reverse refinement module RRM2 is connected to the corresponding input end of the reverse refinement module RRM1; the reverse refinement modules RRM1-RRM5 output prediction maps 1-5 respectively, and the prediction map 1 is a saliency map. Step 4: generating a saliency map using the EGE network: the tensor X and the prediction edge map tensor E are input into the EGE network, after being processed by the encoder, the intermediate layer and the decoder, the prediction maps 1-5 are output, and the prediction map 1 is a saliency map. 2.The method of claim 1, wherein the edge-guided enhancement network is a multi-level edge-guided enhancement network. The information processing procedure of the pre-positioning module PPM is as follows: f'3 = Max(Norm(Conv g (P)))*f3 + f3 where Conv g (·) is a Gaussian convolution kernel, Norm(·) is a normalization process, Max(·) is a maximum value calculation, f3 is a feature tensor output by the feature encoder, and f’3 is a feature tensor output after processing by the predetermined position module PPM module. 3.The method of claim 1, wherein the edge-guided enhancement network is a multi-level edge-guided enhancement network. The intermediate-level adjacent context enhancement module ACEM has the same structure, which includes a local branch, a first adjacent branch, a second adjacent branch, and a fusion module; the local branch is composed of a channel attention module CA and a receptive field module RFB; the first adjacent branch is composed of a downsampling module D and a spatial attention module SA1; the second adjacent branch is composed of an upsampling module U and a second spatial attention module SA2; the fusion module is composed of a first multiplier, a second multiplier, and an adder; the output end of the receptive field module RFB is connected to the input end of the first multiplier, the second multiplier, and the adder, respectively; the input end of the adder is connected to the input end of the local branch, the first multiplier, and the second multiplier, respectively. The first-level and the last-level adjacent context enhancement modules ACEM have different structures from the intermediate-level adjacent context enhancement modules ACEM; the first-level adjacent context enhancement module ACEM includes a local branch and a first adjacent branch, and the local branch is composed of a receptive field module RFB and does not include a channel attention module CA; the last-level adjacent context enhancement module ACEM includes a local branch and a second adjacent branch, and the local branch is composed of a receptive field module RFB and does not include a channel attention module CA. 4.The method of claim 1, wherein the edge-guided enhancement network is a multi-level edge-guided enhancement network. The cross feature fusion module CFF includes convolution kernels CBR1-CBR8. The output data of the convolution kernel CBR1 is multiplied by the output data of the convolution kernel CBR8, and then one path is input into the splicing unit after being processed by the convolution kernel CBR7, and the other path is input into the splicing unit after being processed by the convolution kernels CBR1-CBR3 in turn. The output of the convolution kernel CBR3 is input into the splicing unit after being processed by the convolution kernel CBR8. The output of the splicing unit is added to the output obtained by multiplying the output after being processed by the convolution kernel CBR6 with the predicted edge map tensor E. The output is fused features The convolution kernels CBR1-CBR6 have the same structure, which is composed of a 3*3 convolution layer, a batch normalization layer and a ReLU activation layer. The dilation rate of the convolution kernel CBR1 is 1. The convolution kernel CBR7 is different from the convolution kernel CBR1 in that the dilation rate is 2. The convolution kernel CBR8 is different from the convolution kernel CBR1 in that the dilation rate is 3. 5.The method of claim 1, wherein the edge-guided enhancement network is a multi-level edge-guided enhancement network. The reverse refinement module RRM consists of three branches, the output prediction map p t is obtained by element-wise summation of the three branches; wherein, is the feature map output by ACEM, Conv(·) is a convolution layer with a kernel size of 1x1, is the output of the second branch, is the output of the third branch, denotes element-wise summation; The first branch is an input next-level prediction map. The second branch is expressed as: Y = F (X) + X. The third branch is expressed as: Y = F (X) - X. where CBRG represents three CBR convolutional layers, the first two with a kernel size of 3x3 and a compression of 1, and the last one with a kernel of 1x1, p i denotes the prediction map generated by the current stage, and σ(·) denotes a sigmoid activation function, denotes the output tensor.

Images