Mitigation of face image mismatches via deep feature alignment for robust face recognition

Abstract

Thanks to the advances in deep learning technology, deep learning-based face recognition (FR) has been actively studied, and it shows significantly high performance for various FR benchmark datasets. Accordingly, FR technology has been known as a highly mature technology, and its application to various real-world scenarios is being discussed. However, when the deep learning-based FR algorithm is applied to real-world applications, its performance has been known to be still unsatisfactory. This is mainly attributed to the discrepancy of appearance between training and testing face images. In other words, a face image used for training is usually high-quality, but a face image used for testing is degraded by shooting time, person’s motion, shooting distance between a person and camera, which is referred to as face image mismatch problem. In order to resolve the face image mismatch problem for robust FR, we address the deep feature alignment-based FR algorithm in this dissertation. In particular, we focus on two face image mismatch factors as crucial issues that need to be solved for practical FR: 1) face alignment mismatch (face misalignment) and 2) face image resolution mismatch. In order to deal with the face misalignment problem, we propose a face shape-guided feature alignment learning framework. To tackle the face image resolution mismatch problem, we propose a face image quality-based feature adaptation network between the high-resolution face image, and the low-resolution face image generated by the realistic low-resolution face image generator. Finally, we design a baseline face image quality assessment framework for quantifying the face image mismatch. Through the comparative experiments, we validate the effectiveness of the proposed method under the real-world scenario with the face image mismatch.