A two-stage progressive shadow removal network

Abstract

Removing image shadows has been a challenging task in computer vision due to its diversity and complexity. Shadow removal techniques have been greatly enhanced by deep learning and shadow image datasets, but state-of-the-art methods generally consider the information of the shadow and its neighborhood, ignoring the correlation of the features between the shadow and non-shadow regions. It leads to the resulting image presenting poor overall consistency and unnatural boundary between the original shadow and non-shadow areas. To obtain a consistent and natural shadow removal result, a two-stage progressive shadow removal network is proposed. The first stage performs a multi-exposure fusion network (MEFN) to roughly recover the shadow region features, while in the second stage, a fine-recovery network (FRN) is performed to extract the correlation among the global image contexts, accompanied by a detail feature fusion step. This coarse-to-fine process improves the overall effect of shadow removal, in terms of image quality and boundary consistency. Extensive experiments on the widely used ISTD, ISTD+ and SRD datasets show that the proposed shadow removal network outperforms most of the state-of-the-art methods.