PSR-GAN: Unsupervised Portrait Shadow Removal Using Evolutionary Computing

Because of unwanted occlusion and bad lighting conditions, portrait photographs often suffer from shadows, the presence of which in an image can both decrease its aesthetic quality and increase the difficulty of performing high-level vision tasks. Since most of the shadow removal methods do not specifically remove portrait shadows and hardly delve into the face characteristics, these methods cannot achieve perfect results when removing portrait shadows. Inspired by evolutionary computing, in this paper, we propose a novel unsupervised portrait shadow removal framework, PSR-GAN. To make good use of face characteristics, we introduce a face extraction module, namely FEM, in which we utilize a network to obtain the portrait matte, thereby allowing the network to focus more on the face regions and ignore the interference of background redundant information. Experiments on our collected dataset show that our method is able to effectively remove portrait shadows, and outperforms other existing shadow removal methods.