Scholarlink

<p>Recently, various universal image restoration methods have achieved notable success, employing a single model to handle distinct unknown degradations. However, most of these methods require a significant number of model parameters and computational resources to learn the correlations among multiple degradations. Inspired by the efficient parameter learning methods based on wavelet transform, this paper designs a simple wavelet-based guided transformer model for efficient universal image restoration. First, we utilize wavelet subbands to guide up-and-down sampling and frequency-aware degradation-specific feature aggregation. Then, we send the decomposed high-frequency components to a classifier designed for getting degradation weights. These weights can be used to guide the filtering of detail coefficients along with high-frequency components. Finally, we adaptively aggregate the processed high- and low-frequency components, effectively producing degradation-specific information for universal image restoration. Extensive experiments show the superior performance of the proposed WaveUIR, achieving state-of-the-art (SOTA) performance with 88.16% decrease in model parameters and 77.15% reduction in computational resources compared to previous SOTA method. The code is available at <a href="https://github.com/Archaic-Atom/WaveUIR">https://github.com/Archaic-Atom/WaveUIR</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

WaveUIR: wavelet-based guided transformer model for efficient universal image restoration

  • 期刊论文

  • 发表日期：

    2025年4月11日

查看全文
  • Zhidong Zhu,
  • Bangshu Xiong,
  • Zhibo Rao,
  • Jinhao Zhu,
  • Wenchao Li,
  • Qiaofeng Ou,
  • Xing Li

摘要

Recently, various universal image restoration methods have achieved notable success, employing a single model to handle distinct unknown degradations. However, most of these methods require a significant number of model parameters and computational resources to learn the correlations among multiple degradations. Inspired by the efficient parameter learning methods based on wavelet transform, this paper designs a simple wavelet-based guided transformer model for efficient universal image restoration. First, we utilize wavelet subbands to guide up-and-down sampling and frequency-aware degradation-specific feature aggregation. Then, we send the decomposed high-frequency components to a classifier designed for getting degradation weights. These weights can be used to guide the filtering of detail coefficients along with high-frequency components. Finally, we adaptively aggregate the processed high- and low-frequency components, effectively producing degradation-specific information for universal image restoration. Extensive experiments show the superior performance of the proposed WaveUIR, achieving state-of-the-art (SOTA) performance with 88.16% decrease in model parameters and 77.15% reduction in computational resources compared to previous SOTA method. The code is available at https://github.com/Archaic-Atom/WaveUIR.