Face Description with Local Invariant Features: Application to Face Recognition

Automatic face recognition system based on local feature detection and feature extraction techniques is presented. The method works on color face images and performs face localization initially. It then detects and selects important fiducial facial points and characterizes them by bank of Gabor filters (jets). A well known PCA technique is used to reduce the dimensionality of jets and recognition is realized by measuring the similarity between different jets in eigenspace. A complete investigation on the proposed system is conducted, which covers face recognition under pose, illumination and expression variations of the subjects. The performance of the proposed system is compared with standard methods and it shows the superiority of the proposed system. The proposed approach provides excellent performance by using only 30 fiducial facial feature points and suggests that L1 distance metric is most suitable to measure image similarity in eigenspace. It is an attractive finding that proposed local feature based method is most suitable for machine face recognition application because of its robustness to all types of image variations and computational efficiency.