Shadow Removal Through Learning-Based Region Matching and Mapping Function Optimization

Abstract

In this paper, we develop a novel shadow removal technique where the inputs are a single natural image to be restored and its corresponding shadow mask. We first decompose the image by super-pixels and cluster them into several sim-ilar regions. Then we train a random forest model to pre-dict matched pairs between shadow and non-shadow regions. By applying a distribution-based mapping function on the matched pairs, we can relight pixels in those shadow regions. An optimization framework based on half-quadratic splitting (HQS) method is also introduced to further improve the qual-ity of the mapping process. We also design a post-processing stage with a boundary inpainting function to generate bet-ter visual results. Our experiments show that the proposed method can remove shadows effectively and produce high-quality shadow-free images.