Prof. Taixiang Jiang

Southwestern University of Finance and Economics

Speech title：Rain streaks removal: From directional sparse prior to directional deep prior

Abstract:

Rain streaks removal is an important issue in outdoor vision systems and has recently been investigated extensively. In this talk, I will introduce two methods respectively for video rain streaks removal and single image rain streaks removal. Both two methods are based on one intrinsic property of the rain streaks, i.e., the directionality. In the first work, we construct the deraining model by fully considers the discriminative sparsity in the gradient domain of rain streaks and the clean video caused by the directional property of rain streaks. In the second work, we model the generation of rain streaks via a motion blur model and directly infer the motion blur parameters by a convolutional neural network. Then, the motion blur is used to guide the entire deraining convolutional neural network. Experiments on the synthetic data and real-world rainy data illustrate that our methods achieve state-of-the-art.