TSRRNet: two-stage reflection removal network with reflective guidance
Kuanhong Cheng, Juan Du, Jia Li, Kun Qian, Zhixiao Wang · Applied Optics · 2021
Due to the shortage of paired images, the training of reflection removal networks relies heavily on synthesized samples, for which the ground truths of transmission and reflection are both known. But most existing CNN-based models cannot fully utilize the reflection information, which may cause performance limitations. In this paper, our goal is to design a novel, to the best of our knowledge, network that can take the reflection layer to refine the transmission layer. To this end, we propose a two-stage generative-adversarial-network-based network, where the first stage is used to obtain the coarse estimation of transmission and reflection, and the second stage is used to achieve the refinement. In addition, instead of just applying two penalty terms on the two coarse predictions in previous works, we consider the coarse reflection as a soft mask overlapped on the transmission and apply the recently proposed gated convolution into the second stage for further refinement. The network is trained with an adversarial frame using WGAN. The experimental results with benchmark datasets indicate that our method outperforms several state-of-the-art networks.