Evolving Effective Color Features for Improving FRGC Baseline Performance

This paper presents a novel color feature extraction method for face recognition. Firstly, a new color space, LC1C2, consisting of one luminance (L) channel and two chrominance channels (C1,C2) is introduced as a linear transformation of the input RGB color space. The specific transformation from the RGB color space to the LC1C2 color space is then optimized by Genetic Algorithms (GAs) where a fitness function guides the evolution toward higher recognition accuracy. The feasibility of our feature extraction method has been successfully demonstrated using Face Recognition Grand Challenge (FRGC) databases and the Biometric Experimentation Environment (BEE) baseline algorithm. Specifically, when experimenting with the FRGC version 1 experiment #4, the extracted color features achieve 75% and 73% rank-one face recognition rates using the Principal Component Analysis (PCA) and the Fisher Linear Discriminant (FLD) methods, respectively. When using the FRGC version 2 experiment #4, the extracted color features improve the face verification rate (at 0.1% false acceptance rate) of the BEE baseline algorithm from 12% to 32% and 55% using PCA and FLD, respectively.