Scholarlink

<p>Mirror segmentation is a fundamental task in computer vision for scene understanding in reflective environments. Existing methods still suffer from insufficient cross-modal fusion between appearance and geometric cues, limited complementary modeling of spatial- and frequency-domain information, and inadequate multi-scale and fine-grained feature representation. To address these issues, this paper proposes DCFNet, a dual-domain cross-modal fusion network for RGB-D mirror segmentation. Specifically, the Spatial-Aware Interaction Module (SAI) is designed to promote adaptive cross-modal interaction between RGB and depth features, thereby enhancing their complementary representation. The Frequency-Guided Spatial Attention Module (FGSA) performs collaborative modeling in the spatial and frequency domains, enabling more effective integration of global structural cues and local detailed information. In addition, the Multi-Path Feature Refinement Module (MPFR) aggregates local details, multi-scale contextual information, and global structural dependencies to further improve feature discrimination. Extensive ablation studies and comparative experiments on a public benchmark dataset demonstrate the effectiveness of the proposed design. Experimental results show that DCFNet achieves highly competitive performance against existing state-of-the-art methods and yields clear improvements on several key evaluation metrics.</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

DCFNet: dual-domain cross-modal fusion network for RGB-D mirror segmentation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月3日

查看全文
  • Zhuo Li

摘要

Mirror segmentation is a fundamental task in computer vision for scene understanding in reflective environments. Existing methods still suffer from insufficient cross-modal fusion between appearance and geometric cues, limited complementary modeling of spatial- and frequency-domain information, and inadequate multi-scale and fine-grained feature representation. To address these issues, this paper proposes DCFNet, a dual-domain cross-modal fusion network for RGB-D mirror segmentation. Specifically, the Spatial-Aware Interaction Module (SAI) is designed to promote adaptive cross-modal interaction between RGB and depth features, thereby enhancing their complementary representation. The Frequency-Guided Spatial Attention Module (FGSA) performs collaborative modeling in the spatial and frequency domains, enabling more effective integration of global structural cues and local detailed information. In addition, the Multi-Path Feature Refinement Module (MPFR) aggregates local details, multi-scale contextual information, and global structural dependencies to further improve feature discrimination. Extensive ablation studies and comparative experiments on a public benchmark dataset demonstrate the effectiveness of the proposed design. Experimental results show that DCFNet achieves highly competitive performance against existing state-of-the-art methods and yields clear improvements on several key evaluation metrics.