Applying Multiresolution Analysis to Vector Quantization Features for Face Recognition

Abstract

In this paper, an approach of Facial Parts Detection (FPD) followed by the Discrete Wavelet Transform (DWT) in conjunction with Vector Quantization (VQ) algorithm for Facial Recognition (FR) are proposed. The FR system contains two modes: Training, and Classification. The proposed FR modes contain Preprocessing step followed by the Feature Extraction. The Classification mode yields the identification. The FPD detects nose, both eyes, and mouth for each pose in the Preprocessing step. Then, DWT is employed for each part that is detected for feature selection and data reduction. Thereafter, for further compaction and discrimination, the VQ, with the Kekre Fast Codebook Generation (KFCG) initialization method, is employed to form the final model that contains four feature groups per person. The DWT and VQ are utilized to reduce final feature dimensions without affecting discrimination. The recognition accuracy is calculated using the Euclidean distance. The four databases that are utilized to test the performance of the proposed FR system are: Georgia Tech, YALE, FEI, and FERET. The poses in these databases have various illumination conditions, face rotation, facial expressions, etc. The results, from which samples are presented here, of the FR system and other techniques are obtained and then examined using the Cross Validation based on K-fold method. The proposed FR is shown to improve the recognition accuracies while significantly reducing the storage requirements with comparable computational complexity.