Rain Removal By Image Quasi-Sparsity Priors

Yinglong Wang,Shuaicheng Liu,Chen Chen,Dehua Xie,Bing Zeng

DOI: https://doi.org/10.48550/arXiv.1812.08348

2018-12-20

Computer Vision and Pattern Recognition

Abstract:Rain streaks will inevitably be captured by some outdoor vision systems, which lowers the image visual quality and also interferes various computer vision applications. We present a novel rain removal method in this paper, which consists of two steps, i.e., detection of rain streaks and reconstruction of the rain-removed image. An accurate detection of rain streaks determines the quality of the overall performance. To this end, we first detect rain streaks according to pixel intensities, motivated by the observation that rain streaks often possess higher intensities compared to other neighboring image structures. Some mis-detected locations are then refined through a morphological processing and the principal component analysis (PCA) such that only locations corresponding to real rain streaks are retained. In the second step, we separate image gradients into a background layer and a rain streak layer, thanks to the image quasi-sparsity prior, so that a rain image can be decomposed into a background layer and a rain layer. We validate the effectiveness of our method through quantitative and qualitative evaluations. We show that our method can remove rain (even for some relatively bright rain) from images robustly and outperforms some state-of-the-art rain removal algorithms.