RUSID: Robust Uncertainty-aware Single Image Deraining beyond Certainty

Bingcai Wei, Hui Liu, Chuang Qian, Zijian Li, Zijie Meng

Keywords: Image deraining, Visual priors, Uncertainty-modeling, Prompt learning, SAM

TL;DR: Robust single image rain removal by utilizing uncertainty-modeling visual priors and prompt learning.

Abstract: Rainy weather induces rain streaks, blurs details, and reduces contrast, impairing image quality, making single image deraining a classic research topic. However, existing learning-based image restoration methods fail to account for uncertainties in both data and model dimensions, thus being unable to produce satisfactory results. To address this challenge, we introduce a novel framework called the Uncertainty-aware Visual-priors Prompt-interaction Network (UVPNet). UVPNet comprises three key modules: the Distribution-aware Visual Priors Learning (DVPL) module, which aims at data-wise aleatoric uncertainties, the Certainty-Uncertainty Prompt Fusion (CUPF) module, which tackles model-wise epistemic uncertainties, and the Channel Spatial Uncertainty Weighting Block (CSUWB). UVPNet leverages uncertainty modeling through visual semantic and depth priors and distributionally representative prompts by integrating data-wise and model-wise uncertainty learning. To the best of our knowledge, our UVPNet first utilizes uncertainty modeling with visual priors for single image deraining. Extensive experiment results demonstrate that our UVPNet outperforms state-of-the-art methods on both public synthetic datasets and real-world images while maintaining low complexity.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 23075