Optical remote sensing image salient target detection method based on edge guide enhancement network

Abstract

The invention discloses an optical remote sensing image salient target detection method based on an edge guide enhancement network. The method comprises the steps of preprocessing an image, generating a predicted edge image tensor E, establishing an EGE network, and generating a salient image by using the EGE network. According to the method, the edge information is fully utilized, the object boundary is accurately identified by using the lightweight network, and the predicted edge mapping is generated. The processed trunk characteristics are gradually fused between a high level and a low level, the trunk characteristics are concentrated on a target area, and interference sources are eliminated. The fused features are integrated with deeper features, missing target information is explored from the background through a reverse attention mechanism, and missing target details in the background are well extracted.

Description

Technical Field

[0001] The invention relates to a method for detecting salient targets in optical remote sensing images, and in particular to a method for detecting salient targets in optical remote sensing images based on an edge-guided enhancement network, and belongs to the technical field of computer vision. Background Art

[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 and object detection. Existing SOD methods mostly focus on natural scenes and RGB-T and RGB-D images. With the development of earth observation technology, a large number of remote sensing images have been generated. These images are the products of information obtained by various sensors and are the information carriers of remote sensing detection targets. They can distinguish various details such as water bodies and vegetation. Optical remote sensing images (ORSIs) cover a wide area at high altitudes, with complex imaging conditions and contain 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 changes in object diversity, 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 to fuse multi-resolution features, a parallel downstream fusion network densely connected, and a dense attention fluid network to generate shallow cues to guide the generation of high-level feature maps. To improve edge detection, some methods exploit multi-scale deep features through explicit and implicit edge information. Others use edge-guided recurrent localization networks with parallel decoder structures. Although these methods recognize the importance of edge information, they do not effectively combine edge information with backbone features. In addition, they usually rely on the same structure to extract the boundary and structure information of objects. However, object structure and edge extraction are different tasks, each of which is crucial to the effectiveness of SOD. Therefore, processing the object structure and edge extraction separately will help improve the performance of salient target detection in optical remote sensing images. Summary of the invention

[0004] The object of the present invention is to provide a method for detecting salient targets in optical remote sensing images based on an edge-guided enhancement network.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0006] A method for detecting salient targets in optical remote sensing images based on an edge-guided enhancement network comprises the following steps:

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

[0008] Step 2: Input the tensor X into the edge generation network EGEUNet to generate the predicted edge graph tensor E;

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

[0010] The pre-positioning module PPM is located after the designated feature encoder and is cascaded with each feature encoder in sequence. The output end of each feature encoder and the adjacent feature encoder is 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 final adjacent context enhancement module ACEM outputs the final prediction map through the Conv module; the reverse refinement module RRM receives the output of the next level prediction map and the current level cross feature fusion module CFF to generate the current level prediction map; the first level prediction map is output as a saliency map;

[0011] 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, and after being processed by the encoder, intermediate layer, and decoder, the predicted output is Figure 1 -predict Figure 5 ,predict Figure 1 A significant graph.

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

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

[0014] Among them, Conv g (·) is the Gaussian convolution kernel, Norm(·) is the normalization process, and Max(·) is the 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 module.

[0015] Furthermore, each adjacent context enhancement module ACEM at the intermediate level has the same structure, including a local branch, a first to second adjacent branches and a fusion module; the local branch is composed of a cascaded channel attention module CA and a receptive field module RFB module; the first adjacent branch is composed of a cascaded downsampling module D and a spatial attention module SA1; the first 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 to second multiplier and an adder; the output end of the receptive field module RFB is respectively connected to the input end of the first to second multiplier and the input end of the adder, and the input end of the adder is respectively connected to the input end of the local branch and the input end of the first to second multiplier;

[0016] The difference between the first-level and last-level adjacent context enhancement module structures and the intermediate-level adjacent context enhancement module is that the first-level adjacent context enhancement module includes a local branch and a first adjacent branch, the local branch is composed of an RFB module, and does not include a cascaded channel attention module CA; the last-level adjacent context enhancement module includes a local branch and a second adjacent branch, the local branch is composed of an RFB module, and does not include a cascaded channel attention module CA.

[0017] Furthermore, the cross feature fusion module CFF includes convolution kernels CBR1-CBR8. After the output data of convolution kernel CBR1 and convolution kernel CBR8 are multiplied, one channel is processed by convolution kernel CBR7 and then input into the splicing unit. The other channel is processed by convolution kernels CBR1-CBR3 in sequence and then input into the splicing unit. The output of convolution kernel CBR3 is processed by convolution kernel CBR8 and then input into the splicing unit. The output of the splicing unit is added with the output after multiplication of the predicted edge map tensor E by convolution kernel CBR6, and the fused feature is output. Convolution kernels CBR1-CBR6 have the same structure, consisting of a 3×3 convolution layer, a batch normalization layer and a ReLU activation layer. The expansion rate of convolution kernel CBR1 is 1. The difference between convolution kernel CBR7 and convolution kernel CBR1 is that the expansion rate is 2, and the difference between convolution kernel CBR8 and convolution kernel CBR1 is that the expansion rate is 3.

[0018] Furthermore, the reverse refinement module RRM consists of three branches, and the output prediction map p t It is obtained by element-wise summing of the three branches;

[0019]

[0020] in, is the feature map output by ACEM, Conv(·) is a convolutional layer with a kernel size of 1×1, is the second branch output, is the output of the third branch, ⊕ represents element-wise summation;

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

[0022] The second branch is represented as:

[0023]

[0024] The third branch is expressed as:

[0025]

[0026] Where CBRG represents three CBR convolutional layers, the first two convolution kernels are 3×3 in size, the number of compressed channels is 1, and the last kernel is 1×1, p i represents the prediction graph generated by the current level, σ(·) represents the sigmoid activation function, Represents the output tensor.

[0027] Furthermore, the encoder in the EGE network includes feature encoders RES1-RES5 and a pre-positioning module PPM, the middle layer includes adjacent context enhancement modules ACEM1-ACEM5, and the decoding layer includes a cross feature fusion module CFF1-CFF4 and a reverse refinement module RRM1-RRM4; the feature encoders RES1-RES3, the pre-positioning module PPM, and the feature encoders RES4-RES5 are cascaded in sequence, the input end of the adjacent context enhancement module ACEM1 is respectively connected to the corresponding output ends of the feature encoder RES1 and the feature encoder RES2, and its output end is connected to the corresponding input end of the cross feature fusion module CFF1, and the adjacent context enhancement module The input end of the block ACEM2 is respectively connected to the corresponding output ends of the feature encoder RES1, the feature encoder RES2, and the pre-positioning module PPM, and its output end is connected to the corresponding input end of the cross-feature fusion module CFF2. The input end of the adjacent context enhancement module ACEM3 is respectively connected to the corresponding output ends of the feature encoder RES2, the pre-positioning module PPM and the feature encoder RES4, and its output end is connected to the corresponding input end of the cross-feature fusion module CFF3. The input end of the adjacent context enhancement module ACEM4 is respectively connected to the corresponding output ends of the pre-positioning module PPM, the feature encoder RES4 and the feature encoder RES5, and its output end is connected to the corresponding input end of the cross-feature fusion module CFF4. The corresponding input end, the input end of the adjacent context enhancement module ACEM5 is respectively connected to the corresponding output end of the feature encoder RES4 and the feature encoder RES5, and its output end is connected to the corresponding input end of the cross feature fusion module CFF4, and 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, and 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, and 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. The corresponding input end of the Conv module is connected to 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, and its output end 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, and 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 respectively output the predicted Figure 1 -predict Figure 5 ,predict Figure 1 A significant graph.

[0028] By adopting the above technical solution, the present invention achieves the following technical effects:

[0029] 1. The present invention makes full use of edge information, uses a lightweight network to accurately identify object boundaries and generate predicted edge maps.

[0030] 2. The present invention designs a pre-positioning module (PPM) in the encoder guided by a predicted edge map, which concentrates the main features on the target area and eliminates interference sources.

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

[0032] 4. The present invention uses a cross feature fusion module CFF in the decoder to fuse the characteristics of high-level and low-level layers, and gradually refines the saliency map through reverse attention through the reverse refinement module RRM, so as to better extract the target details missing in the background. BRIEF DESCRIPTION OF THE DRAWINGS

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

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

[0035] Figure 3 It is a structural diagram of the cross feature fusion module CFF of Example 1 of the present invention.

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

[0037] Figure 5 It is the input image of embodiment 1 of the present invention.

[0038] Figure 6 It is a saliency map detected by Example 1 of the present invention. DETAILED DESCRIPTION

[0039] The following examples serve to illustrate the invention.

[0040] Example 1

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

[0042] Step 1: Data preprocessing: Resize the input image to a tensor X of a preset size, input the tensor X into the edge generation network EGEUNet, and generate a predicted edge map tensor E; the edge generation network adopts the technical solution in the document "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 the predicted edge graph tensor E;

[0044] Step 3: Establish an EGE network: The EGE network includes a cascaded encoder, an intermediate layer and a decoder. The encoder includes a feature encoder RES1-RES5 and a pre-positioning module PPM. The intermediate layer includes an adjacent context enhancement module ACEM1-ACEM5. The decoding layer includes a cross feature fusion module CFF1-CFF4 and a reverse refinement module RRM1-RRM4. The feature encoder RES1-RES3, the pre-positioning module PPM, and the feature encoder RES4-RES5 are cascaded in sequence. The input end of the adjacent context enhancement module ACEM1 is connected to the corresponding output end of the feature encoder RES1 and the feature encoder RES2 respectively, and its output end is connected to the cross feature fusion module CFF1-CFF4. The corresponding input end of the adjacent context enhancement module ACEM2 is respectively connected to the corresponding output ends of the feature encoder RES1, the feature encoder RES2, and the pre-positioning module PPM, and its output end is connected to the corresponding input end of the cross feature fusion module CFF2. The input end of the adjacent context enhancement module ACEM3 is respectively connected to the corresponding output ends of the feature encoder RES2, the pre-positioning module PPM and the feature encoder RES4, and its output end is connected to the corresponding input end of the cross feature fusion module CFF3. The input end of the adjacent context enhancement module ACEM4 is respectively connected to the corresponding output ends of the pre-positioning module PPM, the feature encoder RES4 and the feature encoder RES5, and its output end is connected to the cross feature fusion module CFF3. The corresponding input end of the cross feature fusion module CFF4, the input end of the adjacent context enhancement module ACEM5 is respectively connected to the corresponding output end of the feature encoder RES4 and the feature encoder RES5, and its output end 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, and the corresponding output end of the cross feature fusion module CFF2 is respectively connected to the cross feature fusion module The corresponding input end of 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, and its output end 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, and 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 respectively output the predicted Figure 1 -predict Figure 5 .

[0045] The feature encoders RES1-RES5 output the first to fifth features respectively, and the pre-positioning module PPM uses the predicted edge map P as an attention cue to refine the third feature f3. The pre-positioning module PPM aims to emphasize the location of the salient target object and suppress irrelevant information. The information processing process is expressed as:

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

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

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

[0049] The adjacent context enhancement modules ACEM1 and ACEM5 have the same structure. The difference from the adjacent context enhancement modules ACEM2-ACEM4 is that the local branch is composed of RFB modules and does not include the 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 module RFB can perceive multi-scale objects from different receptive fields while reducing parameters and increasing nonlinear depth. The implementation method can be found 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 to reduce the number of channels.

[0052] Since the features at this level are different in scale from the two adjacent features and have different locations of interest in space, adjacent branches process adjacent features. After downsampling and upsampling the left and right branches respectively, the attention maps of two adjacent levels are obtained through the spatial attention module SA, and then the two attention maps are element-wise multiplied with the local branches to cross-align the salient areas.

[0053] The cross feature fusion module CCF includes convolution kernels CBR1-CBR8. After the output data of convolution kernel CBR1 and convolution kernel CBR8 are multiplied, one channel is processed by convolution kernel CBR7 and then input into the splicing unit. The other channel is processed by convolution kernels CBR1-CBR3 in sequence and then input into the splicing unit. The output of convolution kernel CBR3 is processed by convolution kernel CBR8 and then input into the splicing unit. The output of the splicing unit is added with the output after multiplication of the predicted edge map tensor E by convolution kernel CBR6, and the fused feature is output. Convolution kernels CBR1-CBR6 have the same structure, consisting of a 3×3 convolution layer, a batch normalization layer and a ReLU activation layer. The dilation rate of convolution kernel CBR1 is 1. The difference between convolution kernel CBR7 and convolution kernel CBR1 is that the dilation rate is 2, and the difference between convolution kernel CBR8 and 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 target clues missed in the background. RRM consists of three branches. The resulting prediction map p t It is obtained by element-wise summing of the three branches.

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

[0056]

[0057] in, is the feature map output by ACEM, Conv(·) is a convolutional layer with a kernel size of 1×1, is the second branch output, is the output of the third branch, and ⊕ represents element-wise summation.

[0058] The second branch of RRM first connects the prediction map of the previous layer with the current feature. The combined feature map is then processed through cascaded convolutional layers to integrate deeper features and gradually correct and refine the feature map. This process can be expressed as:

[0059]

[0060] The third branch implements reverse attention by leveraging 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 through cascaded convolutional layers, allowing for deeper exploration of 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 convolution kernels are 3×3, the number of compressed channels is 1, and the last kernel is 1×1. i represents the prediction graph generated at each stage, σ(·) represents the sigmoid activation function, Represents the output tensor in a 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, and after being processed by the encoder, intermediate layer, and decoder, the predicted output is Figure 1 -predict Figure 5 ,predict Figure 1 A significant graph.

[0064] In this embodiment, the input image size is adjusted to a tensor X of a preset size, and the tensor size in this embodiment is 3×224×224; after being processed by the adjacent context enhancement modules ACEM1-ACEM5, the number of feature channels is compressed to 64, and the sizes of the first to fifth feature tensors are 64×224×224, 64×112×112, 64×56×56, 64×28×28, and 64×14×14, respectively. The sizes of the first to fifth prediction maps p1-p5 are 1×224×224, 1×112×112, 1×56×56, 1×28×28, and 1×14×14, respectively; in the final result map, we use the first layer prediction map as the saliency map output.

[0065] It should be noted that, at present, a small-scale application survey has been conducted on the technical solution of the present invention, and the survey results show that user satisfaction is relatively high. We have now begun to prepare for technology transformation and application, and have also carried out an intellectual property risk warning survey.

Claims

1. A method for detecting salient targets in optical remote sensing images based on edge-guided enhancement network, characterized in that: The following steps are involved: Step 1: Image preprocessing: resize the input image into a tensor X of a preset size; Step 2: Input the tensor X into the edge generation network EGEUNet to generate the predicted edge graph tensor E; Step 3: Establish an EGE network: The EGE network includes a cascaded encoder, an intermediate layer and a decoder. The encoder includes more than one feature encoder and a pre-positioning module. The intermediate layer includes the same number of adjacent context enhancement modules as the feature encoder. The decoding layer includes a cross feature fusion module and a reverse refinement module. The number of cross-feature fusion modules and reverse refinement modules is the same, which is one less than the number of feature encoders; The pre-positioning module PPM is located after the designated feature encoder and is cascaded with each feature encoder in sequence. The output end of each feature encoder and the adjacent feature encoder is 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 final adjacent context enhancement module ACEM outputs the final prediction map through the Conv module; the reverse refinement module RRM receives the output of the next level prediction map and the current level cross feature fusion module CFF to generate the current level prediction map; the first level prediction map is output as a saliency map; 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. After being processed by the encoder, intermediate layer, and decoder, predicted maps 1 to 5 are output. Prediction map 1 is the saliency map.

2. The method for detecting salient targets in optical remote sensing images based on edge-guided enhancement network according to claim 1, characterized in that: The information processing process of the pre-positioning module PPM is as follows: f’3=Max(Norm(Conv g (P)))*f3+f3 Among them, Conv g (·) is the Gaussian convolution kernel, Norm(·) is the normalization process, and Max(·) is the 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 module.

3. The method for detecting salient targets in optical remote sensing images based on edge-guided enhancement network according to claim 1, characterized in that: Each adjacent context enhancement module ACEM at the intermediate level has the same structure, including a local branch, a first to second adjacent branches and a fusion module; the local branch is composed of a cascaded channel attention module CA and a receptive field module RFB module; the first adjacent branch is composed of a cascaded downsampling module D and a spatial attention module SA1; the first 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 to second multiplier and an adder; the output end of the receptive field module RFB is respectively connected to the input end of the first to second multipliers and the input end of the adder, and the input end of the adder is respectively connected to the input end of the local branch and the input end of the first to second multipliers; The difference between the first-level and last-level adjacent context enhancement module structures and the intermediate-level adjacent context enhancement module is that the first-level adjacent context enhancement module includes a local branch and a first adjacent branch, the local branch is composed of an RFB module, and does not include a cascaded channel attention module CA; the last-level adjacent context enhancement module includes a local branch and a second adjacent branch, the local branch is composed of an RFB module, and does not include a cascaded channel attention module CA.

4. The method for detecting salient targets in optical remote sensing images based on edge-guided enhancement network according to claim 1, characterized in that: The cross feature fusion module CFF includes convolution kernels CBR1-CBR8. After the output data of convolution kernel CBR1 and convolution kernel CBR8 are multiplied, one channel is processed by convolution kernel CBR7 and then input into the splicing unit. The other channel is processed by convolution kernels CBR1-CBR3 in sequence and then input into the splicing unit. The output of convolution kernel CBR3 is processed by convolution kernel CBR8 and then input into the splicing unit. The output of the splicing unit is added with the output after multiplication of the predicted edge map tensor E by convolution kernel CBR6, and the fused feature is output. Convolution kernels CBR1-CBR6 have the same structure, consisting of a 3×3 convolution layer, a batch normalization layer and a ReLU activation layer. The dilation rate of convolution kernel CBR1 is 1. The difference between convolution kernel CBR7 and convolution kernel CBR1 is that the dilation rate is 2, and the difference between convolution kernel CBR8 and convolution kernel CBR1 is that the dilation rate is 3.

5. The method for detecting salient targets in optical remote sensing images based on edge-guided enhancement network according to claim 1, characterized in that: The reverse refinement module RRM consists of three branches, and the output prediction map p t It is obtained by element-wise summing of the three branches; in, is the feature map output by ACEM, Conv(·) is a convolutional layer with a kernel size of 1×1, is the second branch output, is the third branch output, represents element-wise summation; The first branch is to input the next level prediction graph; The second branch is represented as: The third score is expressed as: Where CBRG represents three CBR convolutional layers, the first two convolution kernels are 3×3 in size, the number of compressed channels is 1, and the last kernel is 1×1, p i represents the prediction graph generated by the current level, σ(·) represents the sigmoid activation function, Represents the output tensor.

6. The method for detecting salient targets in optical remote sensing images based on edge-guided enhancement network according to claim 1, characterized in that: The encoder in the EGE network includes feature encoders RES1-RES5 and a pre-positioning module PPM, the middle layer includes adjacent context enhancement modules ACEM1-ACEM5, and the decoding layer includes a cross feature fusion module CFF1-CFF4 and a reverse refinement module RRM1-RRM4; the feature encoders RES1-RES3, the pre-positioning module PPM, and the feature encoders RES4-RES5 are cascaded in sequence, the input end of the adjacent context enhancement module ACEM1 is connected to the corresponding output ends of the feature encoder RES1 and the feature encoder RES2 respectively, and its output end is connected to the corresponding input end of the cross feature fusion module CFF1, and the adjacent context enhancement module ACEM2 is connected to the corresponding input end of the cross feature fusion module CFF1. The input end is respectively connected to the corresponding output end of the feature encoder RES1, the feature encoder RES2, and the pre-positioning module PPM, and its output end is connected to the corresponding input end of the cross-feature fusion module CFF2. The input end of the adjacent context enhancement module ACEM3 is respectively connected to the corresponding output end of the feature encoder RES2, the pre-positioning module PPM and the feature encoder RES4, and its output end is connected to the corresponding input end of the cross-feature fusion module CFF3. The input end of the adjacent context enhancement module ACEM4 is respectively connected to the corresponding output end of the pre-positioning module PPM, the feature encoder RES4 and the feature encoder RES5, and its output end is connected to the corresponding input end of the cross-feature fusion module CFF4. The input end of the enhancement module ACEM5 is respectively connected to the corresponding output ends of the feature encoder RES4 and the feature encoder RES5, and the output end thereof is respectively connected to the corresponding input end of the cross feature fusion module CFF4, and 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, and 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, and 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 reverse refinement module RRM 2, the corresponding output terminal of the cross feature fusion module CFF1 is connected to the corresponding input terminal of the reverse refinement module RRM1, the input terminal of the Conv module is connected to the output terminal of the adjacent context enhancement module ACEM2, and its output terminal is connected to the corresponding input terminal of the reverse refinement module RRM4, the output terminal of the reverse refinement module RRM4 is connected to the corresponding input terminal of the reverse refinement module RRM3, the output terminal of the reverse refinement module RRM3 is connected to the corresponding input terminal of the reverse refinement module RRM2, and the output terminal of the reverse refinement module RRM2 is connected to the corresponding input terminal of the reverse refinement module RRM1; the reverse refinement modules RRM1-RRM5 respectively output prediction images 1-prediction images 5, and the prediction image 1 is a saliency map.

Images