A Novel Efficient Deep Learning Framework for Facial Inpainting

Abstract

The usage of masks during the pandemic has made identifying criminals using surveillance cameras very difficult. Generating the facial features behind a mask is a type of image inpainting. Current research on image inpainting shows promising results on manually pixelated regular holes/patches but has not been designed to handle the specific case of “unmasking” faces. In this paper we propose a novel, custom U-Net based Convolutional Neural Network to regenerate the face under a mask. Simulation results demonstrate that our proposed framework can achieve more than 97% Structural Similarity Index Measure for different types of facial masks across different faces, irrespective of gender, race or color.