Reflection Removal of Display Images on Eyeglasses Using Deep Learning

In videoconferencing applications, the reflections of display images on eyeglasses represent a problem, as they may cause facial expressions to be lost or information to be leaked. Accordingly, the present study was conducted to investigate the deep-learning-based removal of these reflections. To minimize object displacement in images stemming from the motion of eyeglass wearers, we propose a method that rapidly transitions the display to the next screen. In addition, we propose a method that decomposes training images into color channels and resynthesizes them after the neural network is created, accounting for the reflective characteristics of eyeglasses. In our experiments, the proposed method achieved a peak signal-to-noise ratio of 32.6dB, demonstrating efficient removal of reflections. Therefore, our technique can be used to convey facial expressions of the eyes without introducing a risk of information leakage, even when wearing eyeglasses.