Single Image Reflection Removal with Data Augmentation-Based Training Images

Existing learning-based single-image reflection removal approaches have limitations in their generalization ability to handle real-world reflections due to the limited training datasets, which are crucial in reflection removal. Most existing approaches synthesize reflection as a linear combination of background and reflection targets, which cannot well simulate real-world reflections, which are complex and caused by a diverse range of factors. In this paper, instead of synthesizing reflection with a linear additive formulation, we propose to synthesize images contaminated by reflection using a data augmentation algorithm. This algorithm identifies key feature pixel points in a real image contaminated by reflection, then uses a differential evolution algorithm to find the position of the new key feature pixel points, and finally uses affine transformation to generate an augmented mixture image. Multi-to-one image pairs, which consist of multiple augmented mixture images and one ground truth, are similar to multiple-image approaches, making the reflection removal problem less ill-posed. Experimental results show that our reflection image synthesis model based on the data augmentation algorithm helps existing reflection removal methods to remove reflection for achieving higher quality and cleaner images.