Scholarlink

<p>Convolutional neural networks (CNN) have demonstrated impressive detection capabilities in salient object detection (SOD) tasks. However, due to limitations in the capture of global information and modeling long-range dependencies, the performance of CNN-based methods is restricted. In addition, the clarity of the edges also affects the quality of the final detection results. Improving the edge quality of the saliency results remains an open topic. In this paper, we combine edge generation with the SOD task and propose an edge-aware and transformer-based network (EATNet) for RGB-D SOD. Specifically, EATNet mainly consists of a dual-stream backbone, a cross-modal feature enhancement and fusion module (CFEM), and an edge-guide decoder (EGD). The dual-stream backbone, based on the Swin Transformer, first extracts global features from RGB and depth images and models contextual information. The role of CFEM is to achieve the complementary fusion of two different modal features to enhance the understanding of complex scenes. The EGD comprises multiple multi-scale feature aggregation modules (MFAM) aimed at suppressing noise and generating high-quality saliency maps guided by edges obtained from the edge-aware module (EAM). We have conducted extensive experiments on five public RGB-D SOD datasets, and the experimental results demonstrate that the proposed EATNet performs excellently, outperforming 17 state-of-the-art methods. The code and resources can be found at <a href="https://github.com/Qibear/EATNet">https://github.com/Qibear/EATNet</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

EATNet: edge-aware and transformer-based network for RGB-D salient object detection

  • 期刊论文

  • 发表日期：

    2025年2月24日

查看全文
  • Xu Liu,
  • Chenhua Liu,
  • Xianye Zhou,
  • Guodong Fan

摘要

Convolutional neural networks (CNN) have demonstrated impressive detection capabilities in salient object detection (SOD) tasks. However, due to limitations in the capture of global information and modeling long-range dependencies, the performance of CNN-based methods is restricted. In addition, the clarity of the edges also affects the quality of the final detection results. Improving the edge quality of the saliency results remains an open topic. In this paper, we combine edge generation with the SOD task and propose an edge-aware and transformer-based network (EATNet) for RGB-D SOD. Specifically, EATNet mainly consists of a dual-stream backbone, a cross-modal feature enhancement and fusion module (CFEM), and an edge-guide decoder (EGD). The dual-stream backbone, based on the Swin Transformer, first extracts global features from RGB and depth images and models contextual information. The role of CFEM is to achieve the complementary fusion of two different modal features to enhance the understanding of complex scenes. The EGD comprises multiple multi-scale feature aggregation modules (MFAM) aimed at suppressing noise and generating high-quality saliency maps guided by edges obtained from the edge-aware module (EAM). We have conducted extensive experiments on five public RGB-D SOD datasets, and the experimental results demonstrate that the proposed EATNet performs excellently, outperforming 17 state-of-the-art methods. The code and resources can be found at https://github.com/Qibear/EATNet.