Communications in Information and Systems
Volume 22 (2022)
Low-rank appearance-preserving rain streak removal from single images
Pages: 79 – 102
Rain streak removal (RSR) enables the restoration of images affected by rain, facilitating outdoor vision-based tasks. However, conventional wisdoms lead to image degradations when rain is heavy, while learning-based techniques that learn mappings from specific and synthetic datasets hardly generalize and adapt to realworld scenes with unseen patterns. This paper presents a low-rank appearance-preserving RSR algorithm (LA-RSR) for single images. To fully consider the multi-type real-world rainy images, we for the first time formulate a four-prior based optimization function (FPOF) to ensure the performance of both RSR and preserving intrinsic properties (i.e., low-frequency structures and high-frequency details). FPOF is effectively solved by a two-stage decomposition strategy in an iterative way, in which we utilize low-rank matrix recovery and unidirectional total variation (UTV) in the high-frequency component of the rainy image to better separate the rain streak layer and the detail layer. The detail layer is combined with the low-frequency component to yield the final rain-free image. Qualitative and quantitative results show that our approach consistently outperforms the conventional RSR methods and is comparable to the deep learning based methods, without requiring training.
Received 16 July 2020
Published 7 February 2022