Scholarlink

<p>Salient object detection (SOD) seeks to determine the image regions that most prominently attract human visual attention, and these regions are generally represented using saliency maps. However, current approaches often suffer from inadequate interaction between modalities and overly intricate structural configurations. To overcome these challenges, we propose a Multi-scale Cross-modal Fusion Network (MCFNet) for RGB-D SOD tasks. Specifically, MCFNet adopts a dual-stream architecture and employs the Pyramid Vision Transformer V2 (PVTv2) as the encoder to capture multi-scale RGB and depth representations. For feature integration, we propose a Cross-modal Attention Module (CAM), which leverages multiplication and concatenation branches to enhance cross-modal consistency. Furthermore, a Dual-Branch Enhancement Block (DBEB) is introduced to strengthen the fused features while preserving fine details. Our network emphasizes structured architectural optimization, decoupling semantic alignment from complementary information integration. Comprehensive evaluation on six standard benchmark datasets confirms that MCFNet surpasses 16 state-of-the-art methods across four main metrics, thus proving the superiority and efficiency of the method. The source code is publicly available at <a href="https://github.com/zhou-fighting/MCFNet">https://github.com/zhou-fighting/MCFNet</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

Mcfnet: a multi-scale cross-modal fusion network for RGB-D salient object detection

  • 期刊论文

  • 发表日期：

    2026年7月14日

查看全文
  • Wei Wang,
  • Qi Zhou,
  • Lina Huo,
  • Hongxin Geng,
  • Danfeng Lv

摘要

Salient object detection (SOD) seeks to determine the image regions that most prominently attract human visual attention, and these regions are generally represented using saliency maps. However, current approaches often suffer from inadequate interaction between modalities and overly intricate structural configurations. To overcome these challenges, we propose a Multi-scale Cross-modal Fusion Network (MCFNet) for RGB-D SOD tasks. Specifically, MCFNet adopts a dual-stream architecture and employs the Pyramid Vision Transformer V2 (PVTv2) as the encoder to capture multi-scale RGB and depth representations. For feature integration, we propose a Cross-modal Attention Module (CAM), which leverages multiplication and concatenation branches to enhance cross-modal consistency. Furthermore, a Dual-Branch Enhancement Block (DBEB) is introduced to strengthen the fused features while preserving fine details. Our network emphasizes structured architectural optimization, decoupling semantic alignment from complementary information integration. Comprehensive evaluation on six standard benchmark datasets confirms that MCFNet surpasses 16 state-of-the-art methods across four main metrics, thus proving the superiority and efficiency of the method. The source code is publicly available at https://github.com/zhou-fighting/MCFNet.