Double-branch high-resolution saliency target detection method

Abstract

The invention discloses a double-branch high-resolution saliency target detection method, and belongs to the technical field of saliency target detection in high-resolution images. According to the technical scheme, the method comprises a low-resolution positioning branch used for extracting global semantic features and a high-resolution detail refining branch used for learning boundary features of a salient target; the low-resolution positioning branch processes an input high-resolution image through down-sampling, and outputs multi-level coding features and decoding features; and the high-resolution detail refining branch optimizes a salient target boundary through an edge refining module based on a preliminary prediction result of the low-resolution branch. The method has the beneficial effects that by combining a low-resolution positioning branch and a high-resolution detail optimization branch, the contradiction between loss of details and excessive consumption of computing resources during saliency target detection in a high-resolution image is solved; the double-branch network provided by the invention not only significantly improves the precision and efficiency of high-resolution saliency target detection, but also provides a feasible technical scheme for subsequent related research.

Description

Technical Field

[0001] This invention belongs to the field of salient target detection technology in high-resolution images, and specifically relates to a dual-branch high-resolution salient target detection method. Background Technology

[0002] In recent years, deep learning-based salient object detection technology has made significant progress. Initially, image resolution was mainly limited by the capabilities of imaging devices and display technologies. White images had low sharpness, and the quality of color images was also limited. However, with the advancement of digital photography technology, high-resolution images (HR) have gradually become the trend. Improvements in sensor technology and image processing software have increased image resolution; for example, the advent of CMOS sensors has made images sharper. With the development of television and display devices, 4K and 8K resolutions have gradually become mainstream, leading to an increasing demand for high-resolution images. How to perform salient object detection at high resolution has become a hot topic, with preserving detail while avoiding high computational overhead being the main challenge of the High-Resolution Salient Object Detection (HRSOD) task. For example, electronic devices such as digital cameras capture images at very high resolutions (e.g., 1024×2048 pixels or higher), but most salient detection methods can only handle low-resolution images, and few networks directly process high-resolution input. High-resolution input means consuming a lot of computational resources and GPU memory, so the conflict between the need for high-resolution feature maps and a larger receptive field must be resolved. In addition, a balance between accuracy and efficiency must also be considered.

[0003] To address these issues, researchers have explored various approaches. Traditional methods [1-3] progressively obtain semantic feature maps through downsampling, which reduces image size but leads to a loss of detailed information. Upsampling the saliency map back to its original resolution presents a significant challenge, and processing high-resolution input images in this manner may exacerbate these drawbacks. Most previous methods relied on top-level features for prediction, but single-scale semantic features are insufficient to produce satisfactory saliency maps. Low-level features contain weaker semantic information but accurate target locations, while high-level features have richer semantic information but weaker boundary details. Recent methods, such as edge detection, instance segmentation, and semantic segmentation, have achieved some success in enhancing local detail representation, but still struggle to model long-range spatial dependencies, resulting in problems such as blurred boundaries, false positives, and false negatives in saliency prediction in complex scenes. To address the blurred boundary problem, some methods attempt to utilize additional edge information for guidance; however, due to the inherent instability of edge detection, these methods still struggle to maintain accurate saliency boundaries in high-resolution images. Summary of the Invention

[0004] The main objective of this invention is to overcome the shortcomings of existing technologies. This invention provides a dual-branch high-resolution salient target detection method, proposing a new network composed of a low-resolution localization branch and a high-resolution detail refinement branch [4-6]. The low-resolution localization branch enhances the semi-supervised method used in traditional SOD tasks to adapt to HRSOD tasks, introducing the idea of ​​supervised evolution with progressive constraints, providing different supervision labels for each layer. The core operation of this part involves gradually expanding the original ground truth through max pooling to obtain different levels of supervision information, thereby generating rich global semantic features with a larger receptive field. The high-resolution detail refinement branch introduces an edge refinement module, which directly utilizes edge images to reduce noise and fine-tune the boundaries of salient targets, thereby improving detection performance.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A dual-branch high-resolution salient target detection method includes: a low-resolution localization branch for extracting global semantic features and a high-resolution detail refinement branch for learning the boundary features of salient targets.

[0007] The low-resolution localization branch downsamples the input high-resolution image and outputs multi-level encoded and decoded features.

[0008] The high-resolution detail refinement branch is based on the preliminary prediction results of the low-resolution branch, and optimizes the boundaries of significant targets through the edge refinement module.

[0009] Furthermore, the low-resolution localization branch employs a supervised evolutionary mechanism, specifically including: gradually expanding the original ground truth labels through max pooling operations to generate supervised labels at different levels:

[0010]

[0011] Among them, G h This represents the ground truth label at full resolution, down indicates downsampling, and maxpool indicates max pooling.

[0012] Furthermore, the supervised evolutionary mechanism combines evolutionary structure to achieve feature evolution: the upsampled saliency prediction map is generated through a spatial attention mechanism. With fusion features Combination:

[0013]

[0014] Wherein, Fusion represents different levels of FFM, CAM, and BRM in CTDNet, and Indicates pixel-by-pixel multiplication. Indicates decoder features, This represents encoder features.

[0015] Furthermore, the evolutionary structure calculates the evolutionary change graph by constraining feature differences through evolutionary change loss:

[0016]

[0017] in, This represents the evolutionary change graph of the i-th layer. `conv` is a convolutional layer with a kernel size of 3×3. As a supervisor.

[0018] Furthermore, the high-resolution detail refinement branch includes an edge refinement module: based on the initial saliency map P f Given an input image I, an edge map E is generated through multi-scale edge extraction. i :

[0019] E i =s(F2(F1(Cat(I,P) f (6)

[0020] F1 represents the information extraction operation, which consists of three CBR(·) blocks, each of which outputs 64 channels; F2 represents the prediction operation, which consists of one CBR(·) block and one 1×1 convolutional layer.

[0021] Furthermore, the edge refinement module outputs a fused saliency map:

[0022] P final =s(CBR(F1(cat(P) f E i )))) (7)

[0023] Among them, P f For the initial saliency map, E i This is a multi-scale edge map.

[0024] Furthermore, it also includes a feature fusion module, which generates a saliency map through adjacent feature alignment and weighted fusion:

[0025]

[0026] Among them, F i and F i-1 Representing adjacent feature layers, CBR(·) consists of convolution, batch normalization, and ReLU activation functions. Cat(·) and Up(·) represent feature concatenation and 2x upsampling operations, respectively. This represents element-wise addition, while P represents element-wise multiplication.final This represents the final prediction chart.

[0027] Furthermore, the preprocessing module for the low-resolution localization branch includes:

[0028] Standard convolutional layers and 1×1 convolutional layers are used to reduce feature loss;

[0029] Feature channels are recovered through 1×1 convolution and linear operations, reducing the number of model parameters.

[0030] Furthermore, the method employs an improved ResNet-50 backbone network: replacing the standard 3×3 convolutions in the fourth and fifth blocks with dilated convolutions with a dilation rate of 2.

[0031] The beneficial effects of this invention are:

[0032] Compared with existing technologies, the dual-branch high-resolution salient object detection method of this invention solves the contradiction between detail loss and excessive computational resource consumption in salient object detection of high-resolution images by combining a low-resolution localization branch and a high-resolution detail optimization branch. The low-resolution branch adopts a supervised evolutionary mechanism to gradually optimize global semantic features; the high-resolution branch introduces an edge refinement module to accurately restore the target boundary details. By integrating multi-scale information through a feature fusion module, the model significantly improves detection accuracy and boundary clarity while reducing computational burden. The dual-branch network proposed in this invention not only significantly improves the accuracy and efficiency of high-resolution salient object detection, but also provides a feasible technical solution for subsequent related research. Attached Figure Description

[0033] To more clearly illustrate the technical solutions of the embodiments of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and detailed embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0034] in:

[0035] Figure 1 This is a model framework diagram of the present invention;

[0036] Figure 2 This is a diagram of the edge refinement module of the present invention;

[0037] Figure 3 This is a diagram of the feature fusion module of the present invention; Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. The following description, in conjunction with the accompanying drawings... Figure 1-3 The bi-branch high-resolution salient target detection method is further explained.

[0039] Example 1

[0040] To address the aforementioned issues, this invention proposes a two-branch approach for implementing the HRSOD task, comprising a low-resolution localization branch and a high-resolution detail refinement branch. The core idea of ​​this invention is to achieve accurate saliency localization and detail awareness while ensuring computational efficiency. The network of this invention has two branches; the network processes the entire image in both branches, thereby avoiding artifacts caused by patch segmentation operations. Specifically, this invention decouples the HRSOD task into a two-branch structure. The network extracts rich semantic features at a lower resolution to reduce computational cost. The other spatial detail path can learn boundary features of salient objects at a higher resolution, guiding the network to refine the obtained coarse saliency map, ultimately producing a refined prediction result. The low-resolution localization branch can extract rich semantic features at a lower resolution, reducing computational cost. The other detail refinement branch can learn detailed boundary features at high resolution, guiding the network to optimize the coarse map, ultimately producing a refined result. Specifically, the preprocessing module consists of standard convolutional layers and 1×1 convolutional layers. This design aims to mitigate feature loss during processing within depthwise separable convolutional layers. Subsequently, this invention uses 1×1 convolution and accompanying linear operations to recover feature channels, thereby significantly reducing the number of model parameters and speeding up model processing.

[0041] Low-resolution positioning branch

[0042] Supervised Evolution

[0043] In this branch, evolutionary mechanisms at the supervision and feature levels are used to guide the model to achieve accurate and detail-preserving salient target localization in low-resolution images in a manner that progresses from simple to difficult and from coarse to fine, such as... Figure 1 As shown. Specifically, considering computational efficiency, the input high-resolution image is first processed... Perform downsampling It is then further fed into a pre-trained backbone network to extract multi-level encoded features. Next, a feature decoder with an evolutionary structure is employed to facilitate progressive feature learning and layer-by-layer saliency prediction supervised by evolutionary labels, where the decoder features and the predicted saliency maps are represented as follows: and

[0044] As the initial stage of the network, the evolutionary stage lays the foundation for subsequent refinement, focusing on the accurate and complete localization of salient targets. Simultaneously, to meet the high-quality detail requirements of high-resolution scenes, preserving detailed regions in the detection results is crucial. To achieve these goals, this invention incorporates the evolutionary mechanism of a combine harvester, following a principle of gradual progression from easy to difficult and from coarse to fine in both supervision and functionality. This invention enhances the traditional semi-supervised SOD task to adapt to the HRSOD task and proposes the concept of progressively constrained supervised evolution, providing different supervision labels for each layer. The core operation in this part involves gradually expanding the original ground truth labels through max pooling operations to obtain different levels of supervision information.

[0045] In SOD tasks, semi-supervised learning has been widely used to constrain feature learning at each layer of the network [7-9], especially in U-Net-based segmentation models. Specifically, existing methods downsample the ground truth labels or upsample the predicted map to achieve resolution-matched supervised learning. This approach is feasible in traditional SOD tasks, but applying the same precise supervision to both high-level and low-level features in HRSOD tasks may be too stringent. On the one hand, as the network deepens, operations such as pooling and downsampling can blur the boundary information in high-level features. Using ground truth labels with clear boundaries to constrain high-level features, as with traditional supervised methods, may be too harsh and hinder learning. On the other hand, detailed regions constitute a relatively small proportion of the overall ground truth labels, making them easily blurred or ignored in high-level features. Once this happens, it is difficult to recover these details in the final output.

[0046] To address this, this invention proposes a supervised evolution strategy aimed at reducing the learning difficulty of detailed content for high-level features by increasing the proportion of detailed regions in high-level supervision. The design inspiration and core idea of ​​this invention lies in providing different supervision labels for each layer, allowing the prediction map of each layer to gradually evolve in shape, and enabling features to naturally progress from simple to complex and from coarse to fine. This method can reduce the difficulty of network training and learning. This invention primarily focuses on detailed regions of the boundaries, rather than the main target, because these parts are the most difficult to learn but also crucial. In implementation, this invention gradually dilutes the original ground truth labels through max pooling to obtain different levels of supervision information, specifically as follows:

[0047]

[0048] Here, G hThis represents the ground truth label at full resolution, `down` indicates the downsampling operation, and `maxpool` indicates the max pooling operation. The label of the i-th layer is denoted as . As can be seen from the formula, the increased proportion of detailed regions in high-level supervisory labels makes it easier to preserve these details in the final result, thus achieving the second-stage detail optimization. The use of these labels facilitates evolutionary learning across layers of features. This requires capturing the approximate range of salient targets in high-level features, without focusing on complex boundary contour information, which enhances the network's fault tolerance. Through progressive decoding, the boundaries of salient targets are gradually sculpted from high to low. In summary, supervised evolution can better guide the network to learn purposefully and systematically, allowing for the localization and segmentation of salient targets while preserving details.

[0049] Evolutionary structure

[0050] The design motivation for feature evolution stems from two aspects. First, traditional multi-stage fusion schemes often treat cross-layer features equally, while in reality, higher-level features aggregate more global information, achieving coarse localization of salient targets. Therefore, during the decoding process, this invention uses higher-level side outputs as spatial attention maps to suppress background noise in shallow features. For example... Figure 1 As shown, this invention utilizes saliency maps as a medium to establish relationships between adjacent decoder layers. Initially, this invention upsamples decoder features... With corresponding encoder features Fusion. Next, the present invention uses the upsampled saliency prediction map from the previous decoder layer. The spatial attention map is used to refine the fused features, thereby obtaining the decoder features of the current layer. This process can be described as follows:

[0051]

[0052] Here, Fusion refers to the different levels of FFM, CAM, and BRM in CTDNet

[10] , while This represents pixel-wise multiplication. Secondly, previous fusion schemes did not explicitly model the differences between features across layers, leading to an unclear learning direction for the model. Considering that constraints applied only to the supervised layer are insufficient for the network to learn the nature of evolution, and considering that high-level features contain rich semantic information and have good internal consistency between targets, the goal of this invention is to clarify the direction of evolution by constraining the regions. This is beneficial for a coarse-to-fine learning process while maintaining this consistency throughout the layer-wise decoding process. This invention activates the common parts within each layer through multiplication, then uses subtraction to obtain the differences in features, constraining these differences through changes in evolutionary labels. In particular, considering that features in each layer have different and noisy background region representations, this invention suppresses background noise by using saliency maps output from the previous layer, thereby eliminating common internal features. The evolutionary changes are then calculated, and the final evolutionary change prediction is obtained through convolution operations. The above process can be represented as:

[0053]

[0054] here, This represents the evolutionary change graph of the i-th layer, where conv is a convolutional layer with a kernel size of 3×3. Therefore, this invention can use... As supervision, it aims to achieve the desired evolutionary direction. This loss can be expressed as:

[0055]

[0056] ι bce It is a binary cross-entropy loss function.

[0057] High-resolution detail refinement branches

[0058] High-level semantic features contain rich semantic information and play a crucial role in segmentation performance, while low-level features are associated with important spatial details and sharp edge information. Although the global localization branch captures most of the semantic information, it also produces blurred boundaries and some lost details. Spatial details in the image are essential for preserving target boundaries. Furthermore, accurately segmenting salient targets with uniform regions and sharp boundaries remains an unsolved problem. Based on this, this invention introduces a detail refinement branch to preserve the spatial boundary information of the input image, compensating for the lack of detailed high-level semantic information.

[0059] Edge refinement module

[0060] Based on the initial predictions from the previous branch, this invention can identify the location of salient targets; however, the boundaries are still not precise enough. To improve the accuracy of these boundaries, this invention proposes an edge refinement module that directly utilizes the edge image to eliminate noise and fine-tune the boundaries of salient targets. Edge features at various scales are directly extracted from the initial saliency map, and the edge refinement module learns the relationship between these edge features and salient regions to generate a refined saliency map. The edge refinement module consists of two parts. The initial input is the saliency map from the previous stage, and different scales (E... i (i = 1, 2, ...) are obtained using different convolution kernels, and the calculation formula is:

[0061] E i =S(F2(F1(Cat(I,P) f (6)

[0062] Here, F1 represents the information extraction operation, which consists of three CBR(·) blocks, each outputting 64 channels. F2 represents the prediction operation, which consists of one CBR(·) block and one 1×1 convolutional layer; the former outputs 64 channels, and the latter outputs 1 channel. To reduce memory usage during model training, this invention sets i to 2. During network training, the generation of two edge maps is supervised by two corresponding edge labels. As shown in the figure, the edge pixels mapped from edge images at different scales are different. Multiple edge images encourage the network to learn edge information from various angles, providing rich edge cues to help refine edges. This invention overlays the edge images with the initial saliency map to obtain a fused saliency map. This process can be represented as:

[0063] P final =s(CBR(F1(cat(P) f E i )))) (7)

[0064] Here, P final This indicates the output saliency plot.

[0065] Feature fusion module

[0066] In salient object detection tasks, decoders typically use fully convolutional network (FCN) structures to directly aggregate features at various scales. This method combines low-level features that capture details with high-level features that encode semantic information. However, this direct aggregation approach can introduce noise, leading to the loss of detailed features and thus affecting segmentation accuracy. To address these issues, a compressed decoder incorporating neighboring feature fusion (FFM) modules is proposed. These modules progressively aggregate neighboring features to generate an initial salient prediction map. Figure 3As shown, the FFM proposed in this invention is designed to effectively merge adjacent features while preserving important low-level detail features and high-level semantic features. This module selectively eliminates unnecessary features while retaining information-rich features. Specifically, when presenting two adjacent features F... i and F i-1 At that time, the present invention reduces F i To align F with size i-1 The size of F. This invention utilizes F i and F i-1 Element-wise multiplication is performed between them to identify similar features. Then, these similar features are added element-wise to F respectively. i and F i-1 Above. This invention uses global average merging, followed by 1×1 convolution and softmax activation to derive similarity weights. Finally, these similarity weights are compared with F... i and F i-1 Multiply to produce weighted eigenvalues. To restore the eigenvalues ​​to F... i Given the original size, this invention applies a 3×3 convolution operation. This can be represented as:

[0067]

[0068] Here, F i and F i-1 Representing adjacent feature layers, CBR(·) consists of convolution, batch normalization, and ReLU activation functions. Cat(·) and Up(·) represent feature concatenation and 2x upsampling operations, respectively. This represents element-wise addition, while This indicates element-wise multiplication. P find This represents the final prediction chart.

[0069] Example 2

[0070] This invention uses high-resolution datasets (UHRSD, HRSOD) and low-resolution datasets (DUTS, ECSSD) for model training and testing, both of which contain complex scenes. Compared with various methods, this invention, through joint training on multi-scale data, demonstrates significant advantages on both high-resolution and low-resolution datasets, significantly improving the model's generalization ability in both high- and low-resolution scenes.

[0071] This invention employs an improved ResNet-50

[11] . This invention modifies the fourth and fifth blocks of the ResNet-50 architecture. Specifically, all 3×3 standard convolutions in the fourth and fifth blocks are replaced with dilated convolutions of 3×3 with a dilation rate of 2. This modification is to avoid continuous convolution operations that lead to the gradual loss of details in the feature map, while allowing the feature map to maintain a larger receptive field without sacrificing high resolution. In common scenarios such as multiple objects, single objects, complex structures, and low contrast between salient objects and backgrounds, the model consistently performs better in salient object detection, and the output results are very close to the label (GT) in various scenarios. In particular, the model is able to effectively capture the overall structural information of the target object and accurately preserve boundary detail features. However, the method of this invention can significantly improve segmentation accuracy because it can capture complex edge details. These visual comparison results well demonstrate the excellent performance and stability of the method of this invention in challenging scenarios. This advantage stems from three core design innovations: First, the hybrid receptive field convolution module achieves effective integration of local gradient features and global contextual information. Second, the dynamic edge weight allocation strategy enhances the model's feature response capability to fine-grained structures. Third, the cross-scale feature correction network gradually reduces prediction errors through repeated optimization. These findings validate the reliability of the quantification metrics and visually demonstrate the significant progress this method makes in maintaining the integrity of the target structure and reconstructing details, providing crucial technical support for the practical application of salient target detection at high resolution.

[0072] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0073] References

[0074] [1]Zhang PP,Wang D,Lu HC,Wang HY,Ruan X(2017)Amulet: Aggregatingmulti-level convolutional features for salient object detection.Proceedingsofthe IEEE International Conference on ComputerVision,pp 202–211

[0075] [2]Hou QB,Cheng MM,Hu XW,Borji A,Torr P(2017)Deeply supervisedsalient object detection with short connections.2017IEEE Conference onComputer Vision andPattern Recognition(CVPR),pp 3203–3212

[0076] [3]Luo ZM,Mishra A,Achkar A,Eichel J,Marc P(2017)Non-local deepfeatures for salient object detection.2017IEEE Conference on Computer Visionand PatternRecognition(CVPR),pp 6609–6617

[0077] [4]Poudel RP,Liwicki S,Cipolla R(2019)Fast-scnn:fast semanticsegmentation network.arXiv:1902.04502

[0078] [5]Zhao HS,Qi XJ,Shen XY,Shi JP,Jia JY(2018)ICNet for Real-TimeSemantic Segmentation on High-Resolution Images.2018,the European conferenceon computer vision(ECCV),pp 405–420

[0079] [6]Lin GS,Milan A,Shen CH,Reid I(2017)Refinenet:Multi-path refinementnetworks for high-resolution semantic segmentation.Proceedings of the IEEEconference on computer vision andpattern recognition,pp 1925–1934

[0080] [7]Wu,Z.,Su,L.,and Huang,Q.Cascaded partial decoder for fast andaccurate salient object detection.In Proc.CVPR,pp.3907–3916,2019.

[0081] [8]Xie,C.,Xia,C.,Ma,M.,Zhao,Z.,Chen,X.,and Li,J.Pyramid graftingnetwork for one-stage high resolution saliency detection.In Proc.CVPR,pp.11707–11716,2022.

[0082] [9]Liu,Y.,Guo,Q.,Fu,L.,Ke,Z.,Xu,K.,Feng,W.,Tsang,I.W.,and Lau,R.W.H.Structure-informed shadow removal networks.IEEE Trans.Image Process.,32:5823–5836,2023.

[0083]

[10] Zhao,Z.,Xia,C.,Xie,C.,and Li,J.Complementary trilateral decoderfor fast and accurate salient object detection.In Proc.ACM MM,pp.4967–4975,2021.

[11] He K,ZhangX,Ren S,et al.Deep Residual Learning for Image Recognition[C / OL] / / 2016 IEEE Conference on Computer Vision and Pattern Recognition(CVPR),Las Vegas,NV,USA.2016。

Claims

1. A dual-branch high-resolution salient target detection method, characterized in that, include: A low-resolution localization branch is used to extract global semantic features, and a high-resolution detail refinement branch is used to learn the boundary features of salient objects. The low-resolution localization branch downsamples the input high-resolution image and outputs multi-level encoded and decoded features. The high-resolution detail refinement branch is based on the preliminary prediction results of the low-resolution branch, and optimizes the boundaries of significant targets through the edge refinement module.

2. The dual-branch high-resolution salient target detection method as described in claim 1, characterized in that, The low-resolution localization branch employs a supervised evolutionary mechanism, specifically including: gradually expanding the original ground truth labels through max pooling operations to generate supervised labels at different levels: Among them, G h This represents the ground truth label at full resolution, down indicates downsampling, and maxpool indicates max pooling.

3. The dual-branch high-resolution salient target detection method as described in claim 2, characterized in that, The supervised evolutionary mechanism combines evolutionary structure to achieve feature evolution: it uses a spatial attention mechanism to upsample the saliency prediction map. With fusion features Combination: Wherein, Fusion represents different levels of FFM, CAM, and BRM in CTDNet, and Indicates pixel-by-pixel multiplication. Indicates decoder features, This represents encoder characteristics.

4. The dual-branch high-resolution salient target detection method as described in claim 3, characterized in that, The evolutionary structure calculates the evolutionary change graph by constraining feature differences through evolutionary change loss: in, This represents the evolutionary change graph of the i-th layer. `conv` is a convolutional layer with a kernel size of 3×3. As a supervisor.

5. The dual-branch high-resolution salient target detection method as described in claim 1, characterized in that, The high-resolution detail refinement branch includes an edge refinement module: based on the initial saliency map P f Given an input image I, an edge map Ei is generated through multi-scale edge extraction: E i =S(F2(F1(Cat(I,P f )))) (6) F1 represents the information extraction operation, which consists of three CBR(·) blocks, each of which outputs 64 channels; F2 represents the prediction operation, which consists of one CBR(·) block and one 1×1 convolutional layer.

6. The dual-branch high-resolution salient target detection method as described in claim 5, characterized in that, The edge refinement module outputs a fused saliency map: P final =S(CBR(F1(cat(P f ,E i )))) (7) Among them, P f For the initial saliency map, E i This is a multi-scale edge map.

7. The dual-branch high-resolution salient target detection method as described in claim 1, characterized in that, It also includes a feature fusion module, which generates a saliency map through adjacent feature alignment and weighted fusion: Among them, F i and F i-1 Representing adjacent feature layers, CBR(·) consists of convolution, batch normalization, and ReLU activation functions. Cat(·) and Up(·) represent feature concatenation and 2x upsampling operations, respectively. This represents element-wise addition, while P represents element-wise multiplication. final This represents the final prediction chart.

8. The dual-branch high-resolution salient target detection method as described in claim 1, characterized in that, The preprocessing module for the low-resolution localization branch includes: Standard convolutional layers and 1×1 convolutional layers are used to reduce feature loss; Feature channels are recovered through 1×1 convolution and linear operations, reducing the number of model parameters.

9. The dual-branch high-resolution salient target detection method as described in claim 1, characterized in that, The method employs an improved ResNet-50 backbone network: replacing the standard 3×3 convolutions in the fourth and fifth blocks with dilated convolutions with a dilation rate of 2.

Images