Scholarlink

<p>In RGB-T salient object detection (SOD), there exist remarkable discrepancies between the data of the two modalities captured by different sensors. Most of existing studies attempt to integrate multi-modality information through various fusion strategies. However, some of these methods ignore the inherent differences in multi-modality data, while others mostly adopt a uniform fusion strategy when addressing such discrepancies, neglecting that structural alignment and semantic alignment are two distinct yet interdependent hierarchical issues, resulting in poor performance when dealing with some challenging scenarios. In this paper, we explicitly decompose the problem of modality discrepancies into structural alignment and semantic alignment, and propose a Wave-Driven Structure-Semantic Co-Alignment Network (WCANet) for RGB-T SOD. The wave-driven design entails the dual utilization of Wave-MLP as the backbone for feature extraction and wavelet transform as the core mechanism for modality fusion. Specifically, we design a structure-semantic collaborative aligned wavelet fusion module, which relies on two parallel and interactive pathways to address structural misalignment and semantic conflicts, thereby effectively integrating RGB and T modalities. Subsequently, we design a multi-scale progressive grouped feature fusion module to integrate cross-layer information and refine feature representations. Furthermore, we employ knowledge distillation to guide the learning of WCANet and leverage MLP to obtain clear object boundaries for supervising the saliency predictions. Extensive experiments on both RGB-T and RGB-D datasets demonstrate that WCANet outperforms most state-of-the-art methods across various evaluation metrics. Our code is published at <a href="https://github.com/PiPWhiteMoon/WCANet.">https://github.com/PiPWhiteMoon/WCANet.</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

Structure-semantic co-alignment network for RGB-T salient object detection

  • 期刊论文

  • 发表日期：

    2026年4月19日

查看全文
  • Yu Pang,
  • Chuanyue Bai,
  • Lingying Shen,
  • Chengdong Wu

摘要

In RGB-T salient object detection (SOD), there exist remarkable discrepancies between the data of the two modalities captured by different sensors. Most of existing studies attempt to integrate multi-modality information through various fusion strategies. However, some of these methods ignore the inherent differences in multi-modality data, while others mostly adopt a uniform fusion strategy when addressing such discrepancies, neglecting that structural alignment and semantic alignment are two distinct yet interdependent hierarchical issues, resulting in poor performance when dealing with some challenging scenarios. In this paper, we explicitly decompose the problem of modality discrepancies into structural alignment and semantic alignment, and propose a Wave-Driven Structure-Semantic Co-Alignment Network (WCANet) for RGB-T SOD. The wave-driven design entails the dual utilization of Wave-MLP as the backbone for feature extraction and wavelet transform as the core mechanism for modality fusion. Specifically, we design a structure-semantic collaborative aligned wavelet fusion module, which relies on two parallel and interactive pathways to address structural misalignment and semantic conflicts, thereby effectively integrating RGB and T modalities. Subsequently, we design a multi-scale progressive grouped feature fusion module to integrate cross-layer information and refine feature representations. Furthermore, we employ knowledge distillation to guide the learning of WCANet and leverage MLP to obtain clear object boundaries for supervising the saliency predictions. Extensive experiments on both RGB-T and RGB-D datasets demonstrate that WCANet outperforms most state-of-the-art methods across various evaluation metrics. Our code is published at https://github.com/PiPWhiteMoon/WCANet.