Scholarlink

<p>RGB-D salient object detection (SOD) is a fundamental task in the field of computer vision, focusing on identifying the most salient regions within visual inputs by integrating both RGB and depth information. The existing methods primarily transmit encoded features directly to the feature fusion module, which may be impeded by non-essential information contained in these features, thus affecting the predictive accuracy. Besides, few methods explicitly investigate how to refine the features prior to feature transmission stage. To this end, we propose Multi-scale Refinement and Cross-modalities Fusion Network (MRCF-Net), which is comprised of three modules, i.e., inter-layer deep fusion (IDF) module, intra-layer multi-scale refinement (IMR) and multi-modality interaction fusion (MIF). Specifically, the IDF module enhances feature representations by guiding low-level semantics with high-level semantics through an attention mechanism. The IMF incorporates dilated convolution into the Res2Net block, enabling the model to predict the complete structure and clear boundaries of salient objects. Meanwhile, the MIF leverages the complementary and prior relationship between both modalities to capture long-range dependencies for the regulation of cross-modality feature fusion. Extensive experimental results demonstrate that the proposed MRCF-Net achieves superior performance over 27 state-of-the-art RGB-D methods on 5 challenging benchmarks. Our code is publicly available at: <a href="https://github.com/awaitm/MRCF-Net">https://github.com/awaitm/MRCF-Net</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

Red green blue-depth salient object detection based on multi-scale refinement and cross-modalities fusion network

  • 期刊论文

  • 发表日期：

    2025年3月6日

查看全文
  • Kehao Chen,
  • Zhiping Zhou,
  • Kewei Li,
  • Taoyong Su,
  • Zhaozhong Zhang,
  • Jinhua Liu,
  • Chenghao Ying

摘要

RGB-D salient object detection (SOD) is a fundamental task in the field of computer vision, focusing on identifying the most salient regions within visual inputs by integrating both RGB and depth information. The existing methods primarily transmit encoded features directly to the feature fusion module, which may be impeded by non-essential information contained in these features, thus affecting the predictive accuracy. Besides, few methods explicitly investigate how to refine the features prior to feature transmission stage. To this end, we propose Multi-scale Refinement and Cross-modalities Fusion Network (MRCF-Net), which is comprised of three modules, i.e., inter-layer deep fusion (IDF) module, intra-layer multi-scale refinement (IMR) and multi-modality interaction fusion (MIF). Specifically, the IDF module enhances feature representations by guiding low-level semantics with high-level semantics through an attention mechanism. The IMF incorporates dilated convolution into the Res2Net block, enabling the model to predict the complete structure and clear boundaries of salient objects. Meanwhile, the MIF leverages the complementary and prior relationship between both modalities to capture long-range dependencies for the regulation of cross-modality feature fusion. Extensive experimental results demonstrate that the proposed MRCF-Net achieves superior performance over 27 state-of-the-art RGB-D methods on 5 challenging benchmarks. Our code is publicly available at: https://github.com/awaitm/MRCF-Net.