A user dependent multi-resolution approach for biometric data

This paper focuses on the use of a user dependent multi-resolution approach based on local ternary pattern (LTP) in biometric verification. Following an extensive review of the literature on texture descriptors, several methods are compared on well known biometric problems: palm verification and knuckle verification. We propose approaches for extracting a set of local ternary pattern bins separately from the training set of each user, then the Chi square distance is used to compare two templates. The paper is more experimental than novelty in algorithm, our aim is to compare our system with the standard multi-resolution approach, with the novel hierarchical local binary patterns (HLBP) and with different fusions. Extensive experiments conducted over the two well-known biometric characteristics (palmprint and knuckleprint) show the strength of our approach. When each user is given the related selected bins, a near 0 equal error rate is obtained. When the impostor steals the 'selected bins' of the user that he claims to be, our approach slightly outperforms both the standard multi-resolution approach and HLBP. A further improvement in the performance is obtained combining LTP and HLBP.