文献类型：期刊文献

英文题名：Multi-stage feature aggregation transformer for image rain and haze joint removal

作者：Xia, Zhengran[1];Dai, Lei[1];Chen, Zhihua[1];Chen, Kai[1];Li, Ran[1]

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

年份：2025

卷号：149

外文期刊名：ENGINEERING APPLICATIONS OF ARTIFICIAL INTELLIGENCE

收录：;EI(收录号：20251218061913);WOS:【SCI-EXPANDED(收录号:WOS:001451120300001)】；

基金：Acknowledgments This work was supported by the National Natural Science Foun-dation of China (Grant Number: No.62272164 and Grant Number: No.62306113) and the Aeronautical Science Foundation of China (Grant Number: No.202400550S7003 and Grant Number: No. 202400550S7004) .

语种：英文

外文关键词：Image rain-haze joint removal; Multi-stage feature aggregation; Information calibration

摘要：This study investigates the complex meteorological phenomenon of rain-haze mixtures, resulting from water vapor condensation in real-world rainy environments. Current methodologies predominantly focus on rain removal, often neglecting the image degradation caused by haze. We propose a multi-stage feature aggregation transformer (MFAFormer). The model consists of multiple stages, each comprising an information calibration module and a feature aggregation module. Specifically, the information calibration module extracts local high-frequency rain features, whereas the feature aggregation module extracts global low-frequency haze features and integrates features across stages. Benefiting from multi-stage feature extraction and aggregation, MFAFormer can efficiently and robustly address the complex degradations caused by the simultaneous presence of rain and haze. Additionally, we propose a novel rain and haze mixed dataset RainHaze Synscapes. The dataset contains large variations in rain streaks, haze density, and scene contents. Experimental results indicate that the model surpasses other methods on the dataset. MFAFormer demonstrates a remarkable improvement, achieving a 37.46% increase in peak signal-to-noise ratio (PSNR) and a 10.48% increase in structural similarity (SSIM) compared to the baseline model on the RainHaze Synscapes dataset. This accomplishment offers valuable insights for the research of rain and haze joint removal domain.