Encode Imaging System Parameters as Distribution to Improve Reflection Generation and Removal

Images captured behind glass are often contaminated by the undesired reflection that hinders other computer vision tasks. Due to ubiquitous glasses, the removal of undesired reflection becomes more important. Reflection is determined by imaging system parameters such as the transparency, color of the glasses, the position of the camera. In this letter, we propose a novel generative model based on a generative adversarial network (GAN), where an Encoder extracts system parameters from input images and encodes them succinctly as a distribution and a Generator emulates the optical process of reflection. Separating the optical process of the system parameters allows our model to cope with diversified real-life scenarios and avoid the mode collapse phenomenon. Based on the generative model, we introduce a reflection removal model that simultaneously extracts the original image, the reflection layer, and the encoded parameters from single image input. Computational results of real data show that our approach outperforms existing approaches.