Shadow Removal Based on 2Cycles-GAN

Abstract

Shadow removal and restoration of the image content in the shadowed regions by using shadow removal has become more and more popular in computer vision. But almost all current shadow-removal approaches use shadow-free images for training. Recently, an innovative approach that trains samples without this requirement due to this method crops patches with and without shadows from shadow images. However, it is insufficient to directly learn the essential relationships between shadow and shadow-free domains using adversarial learning and cycle-consistency constraints. Moreover, constructing many of these unpaired patches is still time-consuming and laborious. In our paper, we propose a new method named 2Cycles-G2R-ShadowNet. A shadow mask is used in our framework. We use the mask to guide the shadow generation to reformulate cycle-consistency constraints. To weakly-supervised shadow removal, we train shadow images and corresponding masks to leverage shadow generation. In our 2Cycles-G2R-ShadowNet, three subnetworks are used for shadow generation, shadow removal, and image post-processing, and we jointly train and test them end-to-end. Our method can optimize the performance by simultaneously learning to produce shadow masks and remove shadows. Extensive experiments on the ISTD dataset show that 2Cycles-G2R-ShadowNet achieves competitive performances and outperforms the current state of arts and patch-based shadow-removal method.