Region Based Hierarchical Modelling for Effective Shadow Removal in Natural Images

Shadow removal from natural images is always a challenging but intriguing issue. In this paper, we develop a novel region based hierarchical model to address this problem. Firstly, an automatic shadow detection method is applied to obtain the shadow region and then roughly decompose it into umbra and penumbra areas. A hierarchical framework is then adopted to relight umbra and penumbra separately. For umbra arrears, each pixel is recovered via a lighting model on the basis of the characteristics of shadow-free regions. Afterwards, we present an improved brute force filling image in-painting strategy to relight complex penumbra areas using the information of neighboring patches which are selected from recovered umbra and lit regions according to their similarities in texture, gradient and distance. The comparing experimental results with three canonical and other state-of-the-art approaches have demonstrated the superiority and versatility of our method in generating high quality shadow-free images with consistent illumination.