Scholarlink

<p>Many existing stereo image deraining methods do not fully exploit the correlation information between left and right views, as well as the image background prior. This paper introduces a Residual Channel Prior-Driven Interaction Multi-Scale Network (RPIM-Net) for stereo image deraining, comprising feature extraction, parallax interaction, and reconstruction. To enhance deraining performance, we introduce Residual Channel Prior (RCP) to obtain richer background features. Additionally, we design Residual Multi-Path Smooth Dilated convolution Block (ResMSDBlock) to capture multi-scale features while mitigating discontinuities between adjacent pixels caused by the dilated convolution. To further refine the deraining process, we incorporate the bi-directional Parallax Attention Module (biPAM) which leverages correlation between left and right views to enhance view reconstruction. Finally, a channel attention module fuses features after the view interaction, followed by a residual dense block to reconstruct high-quality rain-free images. Experimental results illustrate the superiority of the proposed RPIM-Net algorithm on real-world rain images and synthetic datasets over the existing algorithms. The source code is released at <a href="https://github.com/ooowoooyoooh/RPIM-Net/tree/master/stereoDeraining">https://github.com/ooowoooyoooh/RPIM-Net/tree/master/stereoDeraining</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

RPIM-net: residual channel prior-driven interaction multi-scale network for stereo image deraining

  • 期刊论文

  • 发表日期：

    2025年6月7日

查看全文
  • Yihan Wang,
  • Yongfang Wang,
  • Mengyao Li

摘要

Many existing stereo image deraining methods do not fully exploit the correlation information between left and right views, as well as the image background prior. This paper introduces a Residual Channel Prior-Driven Interaction Multi-Scale Network (RPIM-Net) for stereo image deraining, comprising feature extraction, parallax interaction, and reconstruction. To enhance deraining performance, we introduce Residual Channel Prior (RCP) to obtain richer background features. Additionally, we design Residual Multi-Path Smooth Dilated convolution Block (ResMSDBlock) to capture multi-scale features while mitigating discontinuities between adjacent pixels caused by the dilated convolution. To further refine the deraining process, we incorporate the bi-directional Parallax Attention Module (biPAM) which leverages correlation between left and right views to enhance view reconstruction. Finally, a channel attention module fuses features after the view interaction, followed by a residual dense block to reconstruct high-quality rain-free images. Experimental results illustrate the superiority of the proposed RPIM-Net algorithm on real-world rain images and synthetic datasets over the existing algorithms. The source code is released at https://github.com/ooowoooyoooh/RPIM-Net/tree/master/stereoDeraining.