文献类型：期刊文献

英文题名：Feature-refined adaptive modulation transformer for image deraining

作者：Huang, Yeting[1];Dai, Lei[1];Chen, Zhihua[1];Hu, Wenlong[1];Wang, Shouli[1]

机构：[1]East China Univ Sci & Technol, Dept Comp Sci & Engn, Shanghai 200237, Peoples R China

年份：2025

卷号：157

外文期刊名：ENGINEERING APPLICATIONS OF ARTIFICIAL INTELLIGENCE

收录：;EI(收录号：20252518635754);WOS:【SCI-EXPANDED(收录号:WOS:001518599400005)】；

基金：This work was supported by the National Natural Science Foundation of China (Grant No. 62272164 and No. 62306113) .

语种：英文

外文关键词：Image deraining; Detail guidance; Feature refinement; Frequency modulation

摘要：Recent image deraining methods demonstrate impressive reconstruction performance by leveraging the global modeling capability of Transformer architecture. However, unlike convolutional approach, Transformer inherently struggles to capture high-frequency detail effectively. Furthermore, existing methods primarily focus on spatial information while largely neglecting the frequency-domain characteristics of rain streaks, which are crucial for rain removal. To address these challenges, we propose a feature-refined adaptive modulation Transformer (FRAMT), which effectively integrates spatial-domain features with frequency-domain modulation to enhance deraining performance. To accurately identify rain streaks and efficiently separate them from the background, the detail-guided attention block enhances sensitivity to high-frequency components by integrating pooling operation with convolution. To mitigate image blurring and detail loss induced by rain streaks, the local feature refinement block employs a multi-scale content decomposition strategy, utilizing a parallel multi-branch architecture to extract diverse contextual features across varying spatial scales. Additionally, the adaptive fusion modulation block incorporates a frequency selection mechanism that dynamically modulates feature response, effectively suppressing redundant information and irrelevant features. Extensive experiments conducted on widely used benchmark datasets demonstrate that the proposed method is more competitive than advanced methods.