Finger kunckcle patterns based person recognition via bank of multi-scale binarized statistical texture features

Abstract

This paper proposes a novel finger knuckle patterns (FKP) based biometric recognition system that utilizes multi-scale bank of binarized statistical image features (B-BSIF) due to their improved expressive power. The proposed system learns a set of convolution filters to form different BSIF feature representations. Later, the learnt filters are applied on each FKP traits to determine the top performing BSIF features and respective filters are used to create a bank of features named B-BSIF. In particular, the presented framework, in the first step, extracts the region of interest (ROI) from FKP images. In the second step, the B-BSIF coding method is applied on ROIs to obtain enhanced multi-scale BSIF features characterized by top performing convolution filters. The extracted feature histograms are concatenated in the third step to produce a large feature vector. Then, a dimensionality reduction procedure, based on principal component analysis and linear discriminant analysis techniques (PCA + LDA), is carried out to attain compact feature representation. Finally, nearest neighbor classifier based on the cosine Mahalanobis distance is used to ascertain the identity of the person. Experiments with the publicly available PolyU FKP dataset show that the presented framework outperforms previously-proposed methods and is also able to attain very high accuracy both in identification and verification modes.