BiPR-RL: Portrait relighting via bi-directional consistent deep reinforcement learning

Abstract

Portrait relighting via deep reinforcement learning could conduct portrait relighting by sequentially predicting local light editing strokes, simulating image editing by artists using brush strokes. It is a locally effective, scale-invariant and interpretable method, which yields state-of-the-art results in relighting wild portrait images. However, it still has problems due to the lack of supervision information on intermediate processes, which makes the reinforcement learning agent not receive sufficient feedback during the exploration process. This will cause incorrect strokes in the relighting. To further optimize the relighting effect, we take advantage of the fact that the light editing actions are invertible and the inverse actions can generate a backward sequence of states. The backward state sequence also provide useful supervision signals as they convey information about the future states given the action choices. Therefore, we propose to combine the forward and the backward state sequence predictions to improve the learning efficiency for portrait relighting. Using the bi-directional state sequences, we design corresponding bi-directional consistent rewards to guide the model to explore the actions with higher accuracy to maximize the performance of the proposed method. The proposed approach is used for portrait relighting tasks based on both SH-lighting and reference images. The results show that our method has further improved the performance by avoiding wrong strokes. Meanwhile, our method performs better than SOTA methods in producing locally effective relighting images for wild portrait images.