De-Glared: Eyeglasses Glare and Reflection Removal Using Deep Neural Networks

Eliminating reflections from eyeglasses in digital images presents a formidable challenge necessitating the deployment of deep learning techniques. When photographs featuring individuals wearing eyeglasses are captured, reflections and glare from various angles can obscure the eyes, rendering them indistinct. This impairment adversely affects image quality and poses significant hurdles for computer vision tasks such as eye identification, segmentation, and facial recognition. Our research has devised an effective neural network based on U-Net, proficiently designed to eradicate glare and reflections from digital eyeglass-clad images, thereby yielding crisp and unobstructed eye visuals. In constructing our dataset, we meticulously assembled 2541 images as we encountered a scarcity of extensive datasets containing human images both with and without eyeglass reflections. This study involved a comprehensive analysis of various algorithms and these algorithms underwent rigorous training using this dataset. Our comprehensive analysis unveiled that U-Net exhibited unparalleled performance in contrast to alternative algorithms, boasting an impressive PSNR value of 38.48 dB. Furthermore, it has demonstrated proficiency in effectively mitigating reflection glare in images.