Scholarlink

<p>The task of single-image deraining focuses on generating clear images from rainy ones, a crucial challenge with wide-ranging applications. Although Transformer models have achieved notable success in this field, their quadratic computational complexity hinders practical use. To address this, we propose a lightweight and efficient solution called the Focused Network (FoNet), which is designed to handle varying degrees of degradation across different image regions. FoNet incorporates a Binary Domain Selection Mechanism to emphasize critical features such as edges and highly degraded areas, while Multi-scale Residual Units are used to improve both efficiency and performance. Built on a U-net architecture, the network demonstrates a PSNR improvement of + 0.62&#xa0;dB over state-of-the-art methods, processes images within 9&#xa0;ms, and requires only 3.74&#xa0;million parameters. These findings highlight the practicality and effectiveness of our approach in advancing single-image deraining.</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

FoNet: Focused Network for Single Image Deraining

  • 期刊论文

  • 发表日期：

    2025年2月7日

查看全文
  • Sambasiva Rao Gumma,
  • Balaram Murthy Chintakindi

摘要

The task of single-image deraining focuses on generating clear images from rainy ones, a crucial challenge with wide-ranging applications. Although Transformer models have achieved notable success in this field, their quadratic computational complexity hinders practical use. To address this, we propose a lightweight and efficient solution called the Focused Network (FoNet), which is designed to handle varying degrees of degradation across different image regions. FoNet incorporates a Binary Domain Selection Mechanism to emphasize critical features such as edges and highly degraded areas, while Multi-scale Residual Units are used to improve both efficiency and performance. Built on a U-net architecture, the network demonstrates a PSNR improvement of + 0.62 dB over state-of-the-art methods, processes images within 9 ms, and requires only 3.74 million parameters. These findings highlight the practicality and effectiveness of our approach in advancing single-image deraining.