Scholarlink

<p>In the past few years, the requirements for the quality of nighttime images have been constantly increasing. However, flares can brighten dark areas, alter the frequency characteristics of the image, reduce contrast, and consequently affect the visual quality. In this paper, we present a spatial-frequency synergy transformer-based network, called SFS-Former, which aimed to address these challenges. Specifically, there are two core designs. First, we introduce a multi-scale attention block (MSAB), that performs shifted window-based self-attention and global information extraction based on convolution. It acquires local information while preserving global information. Second, we connect the deep spatial frequency block (DSFB) after the MSAB. This block enables the extraction of global information. Powered by these two designs, the SFS-Former has a high capability for capturing both local and global dependencies for nighttime flare removal. Extensive experiments on Flare7k++ datasets prove the effectiveness and superiority of the proposed SFS-Former against state-of-the-art models both qualitatively and quantitatively.</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

Spatial-frequency synergy transformer for nighttime flare removal

  • 期刊论文

  • 发表日期：

    2025年5月28日

查看全文
  • Shun Zhao,
  • Bo Wang

摘要

In the past few years, the requirements for the quality of nighttime images have been constantly increasing. However, flares can brighten dark areas, alter the frequency characteristics of the image, reduce contrast, and consequently affect the visual quality. In this paper, we present a spatial-frequency synergy transformer-based network, called SFS-Former, which aimed to address these challenges. Specifically, there are two core designs. First, we introduce a multi-scale attention block (MSAB), that performs shifted window-based self-attention and global information extraction based on convolution. It acquires local information while preserving global information. Second, we connect the deep spatial frequency block (DSFB) after the MSAB. This block enables the extraction of global information. Powered by these two designs, the SFS-Former has a high capability for capturing both local and global dependencies for nighttime flare removal. Extensive experiments on Flare7k++ datasets prove the effectiveness and superiority of the proposed SFS-Former against state-of-the-art models both qualitatively and quantitatively.