Scholarlink

<p>In recent years, salient object detection (SOD) methods for RGB-D have achieved many impressive results with depth modality. However, most SOD methods rely on bulky networks with larger parameters and complex structures (including edge enhancement, pseudo-depth estimation, etc) to achieve higher detection accuracy, which hinders the application of RGB-D technology in computer vision tasks. Therefore, how to design an RGB-D SOD model with excellent detection accuracy under the conditions of small parameters and low complexity is an open issue. To address this issue, this paper proposes a Parameter-efficient yet Structure-concise Network for RGB-D SOD, called PSNet, in which a unidirectional guidance strategy is explored to enhance the feature representation capability of the RGB stream. Specifically, PSNet utilizes two lightweight networks as an asymmetric dual-stream encoder to extract features from two modalities, resulting in a model with fewer parameters and lower computational complexity. And, a Cross-modal fusion module (CFM) is given to achieve feature fusion by fully exploring the common information between RGB and depth modalities. Furthermore, a progressive decoder consisting of three Residual Interaction Modules (RIMs) is designed to assist the network in more accurately inferring salient objects. The collaborative work between the encoder, CFM, and decoder can help the model efficiently and accurately detect salient objects. Extensive experimental results demonstrate the superiority of the proposed PSNet over the latest 24 state-of-the-art (SOTA) methods on 6 benchmark datasets. The PSNet has 15.4M Params, 15.1G FLOPs, and 41.5 FPS. The codes and results are available at <a href="https://github.com/hjy0518/PSNet">https://github.com/hjy0518/PSNet</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

PSNet: a Parameter-efficient and structure-concise network for RGB-D salient object detection

  • 期刊论文

  • 发表日期：

    2026年2月9日

查看全文
  • Haopeng Wang,
  • Jinyu Han,
  • Hua Liu,
  • Haojie Li

摘要

In recent years, salient object detection (SOD) methods for RGB-D have achieved many impressive results with depth modality. However, most SOD methods rely on bulky networks with larger parameters and complex structures (including edge enhancement, pseudo-depth estimation, etc) to achieve higher detection accuracy, which hinders the application of RGB-D technology in computer vision tasks. Therefore, how to design an RGB-D SOD model with excellent detection accuracy under the conditions of small parameters and low complexity is an open issue. To address this issue, this paper proposes a Parameter-efficient yet Structure-concise Network for RGB-D SOD, called PSNet, in which a unidirectional guidance strategy is explored to enhance the feature representation capability of the RGB stream. Specifically, PSNet utilizes two lightweight networks as an asymmetric dual-stream encoder to extract features from two modalities, resulting in a model with fewer parameters and lower computational complexity. And, a Cross-modal fusion module (CFM) is given to achieve feature fusion by fully exploring the common information between RGB and depth modalities. Furthermore, a progressive decoder consisting of three Residual Interaction Modules (RIMs) is designed to assist the network in more accurately inferring salient objects. The collaborative work between the encoder, CFM, and decoder can help the model efficiently and accurately detect salient objects. Extensive experimental results demonstrate the superiority of the proposed PSNet over the latest 24 state-of-the-art (SOTA) methods on 6 benchmark datasets. The PSNet has 15.4M Params, 15.1G FLOPs, and 41.5 FPS. The codes and results are available at https://github.com/hjy0518/PSNet.