Shadow removal via dual module network and low error shadow dataset

Abstract

Shadow removal is a challenging task as it requires us to recover common penumbra at the shadow boundary and trys not to alter the illumination of the non-shadow regions. In this work, we propose a novel strategy to achieve these goals based on Generative Adversarial Networks and design a dual-module architecture (DM-GAN) consisting of a mask generator, a matte generator, an umbra module, a penumbra module, and a boundary generator. The mask generator and matte generator first produce a shadow alpha mask and a shadow matte for input shadow image. Combined with this mask and matte, we employ the umbra module and penumbra module to generate an umbra removal image and a mask of the penumbra. Finally, we recover the penumbra based on the idea of image inpainting to obtain a boundary-smoothing result with less alteration at non-shadow regions. Besides, we construct a Low Error Shadow Dataset (LESD) with less error and more scenes to maintain the global illumination consistency between shadow and non-shadow regions. Extensive experiments demonstrate that our proposed DM-GAN can outperform other state-of-the-art methods.