A ZigZag Pattern of Local Extremum Logarithm Difference for Illumination-Invariant and Heterogeneous Face Recognition

Abstract

A novel methodology for matching of illumination-invariant and heterogeneous faces is proposed here. We present a novel image representation called local extremum logarithm difference (LELD). Theoretical analysis proves that LELD is an illumination-invariant edge feature in coarse level. Since edges are invariant in different modalities, more importance is given on edges. Finally, a novel local zigzag binary pattern LZZBP is presented to capture the local variation of LELD, and we call it a zigzag pattern of local extremum logarithm difference (ZZPLELD). For refinement of ZZPLELD, a model based weight value learning is suggested. We tested the proposed methodology on different illumination variations, sketch-photo and NIR-VIS benchmark databases. Rank-1 recognition of 96.93% on CMU-PIE database and 95.81% on Extended Yale B database under varying illumination, show that ZZPLELD is an efficient method for illumination invariant face recognition. In the case of viewed sketches, the rank-1 recognition accuracy of 98.05% is achieved on CUFSF database. In the case of NIR-VIS matching, the rank-1 accuracy of 99.69% is achieved and which is superior to other state-of-the-art methods.