Unifying Domain Adaptation into Simultaneous Discriminant Analysis: A Method for Low-Resolution Face Recognition with Single Sample Per Person

Abstract

In some real-world applications, such as law enforcement, it is required to find out the true identity of an LR face image by using only the high-resolution (HR) profile face image. This leads to LR face recognition with single sample per person (SSPP). As it is intractable to find a match between LR and HR images directly and difficult to explore variations from SSPP, LR face recognition with SSPP is quite a challenging problem. To address the problem, based on simultaneously discriminative analysis (SDA), we propose to introduce an auxiliary dataset (source domain) containing multiple samples per person and integrate domain adaptation into the learning of coupled mappings. In the proposed method, the discriminative analysis of source domain and target domain, together with the domain adaptation between the source domain and target domain are unified into one framework, which produces Enhanced SDA. In this framework, the distribution mismatch is reduced, and the information in both the source domain and target domain are used to learn the coupled mappings. Enhanced SDA is extensively evaluated on LFW and SC face databases, the promising results of Enhanced SDA show its effectiveness in LR face recognition with single sample per person.