Person identification from lip texture analysis

Abstract

The interactive liveness detection for fact recognition often requires users to read some digits from 0 to 9. The movement and variation of lip texture during reading potentially provide discriminative information for human identification. This paper firstly addressed the issue of whether the lip texture during reading can serve as a soft-biometric for person identification. Different from the traditional lip recognition methods that are based on color statistics and lip shapes, we develop a deep architecture that incorporates both CNN and LSTM to jointly model the appearance and the spatial-temporal information of lip texture. We also build a new lip recognition database that contains 11,123 videos for the number 0∼9 in Chinese from 57 people. Experimental results show that the proposed method can achieve 96.01% on close-set protocols, suggesting the usage of lip texture as soft-biometrics for facilitating face recognition.