Deep Variational Inference Network for Single Image Reflection Removal

Abstract

Reflection removal from an image with undesirable reflections is a challenging and ill-posed problem in low-level vision. In recent years, several deep learning approaches have been proposed to tackle the task of single image reflection removal (SIRR). These methods, however, do not fully utilize the fundamental image priors of reflection and lack interpretability. In this paper, we propose a deep variational inference reflection removal (VIRR) method for the SIRR problem, which has good interpretability and good generalization ability. Based on the proposed VIRR method, the posterior distributions of the latent transmission and reflection images can be estimated jointly through variational inference, using deep neural networks. Furthermore, the proposed network framework can be trained by the supervision of data-driven priors for the transmission image and reflection image, which is produced by the variational lower bound objective of marginal data likelihood. Our proposed method outperforms previous state-of-the-art approaches on four benchmark datasets, as demonstrated by extensive subjective and objective evaluations.