Interactive shadow removal using a conditional generative adversarial network

Abstract

We propose an interactive shadow removal method, which uses a multi-branch conditional Generative Adversarial Network. In contrast to previous work that drastically change the whole image or change the color of unwanted area of the shadow image, our method preserves semantic information of shadow which leads to a more natural shadow removed image with only a coarse and simple shadow mask annotated by the user. Our multi-branch network generates accurate shadow mask and shadow free image. With skip connection to shadow free generator from shadow mask generator, the network can accurately point where the image should be edited through the network. Moreover, our two discriminators for shadow mask and shadow free image help both shadow mask and shadow free image generator to model higher level semantic information. Extended experiment with a publically available dataset shows that our method significantly improves in both shadow removal and detection compared our method to other state-of-art.