Perceptual and Structural Similarity-Driven Loss for Improving Unified Rain Removal

View at publisher Summarise with VIRAT
Authors: Yusuke Yamamoto, Yinhao Li, Hiroshi Taga, Koki Iwasa, Ryuichi Shichikawa, Makoto Suganami, Kazuhiro Nakamoto, Yen-Wei Chen
Cited by: 0
References: 11

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Images captured in rainy scenes suffer from various non-uniform degradations—such as streaks, raindrops, and haze—that reduce visibility and impair downstream perception tasks. Most existing approaches train deep models with an L1 loss, which often fails to fully remove rain noise or inadvertently destroys structural details of background objects. In this work, we propose a composite objective that integrates structural and perceptual components, enabling deraining models to achieve both effective noise removal and high-quality visual results aligned with human perception. To validate the effectiveness of our method, we employ a unified Medical Image Restoration model adapted to natural images with rain-haze mixtures and systematically compare different loss functions. Experiments on a mixed rain/haze benchmark demonstrate that our loss consistently improves LPIPS scores and produces visually clearer restorations, while maintaining PSNR/SSIM comparable to strong baseline models.

Fields of Study & Subjects

Image Enhancement Techniques Image and Signal Denoising Methods Advanced Image Processing Techniques

Publication details

Journal / source
International Conference on Consumer Electronics
Pages
1–4
Publication date
3 February 2026
Language
EN