Nonseparable wavelet domain BPCA for face recognition

Face recognition has been studied extensively recently. The main difficulty faced by the current face recognition techniques stems from large variations in facial expression, pose and illumination. This paper presents an effective method for face recognition using nonseparable wavelet domain block-based PCA(BPCA) method. Our investigations demonstrate that the constructed nonseparable wavelet can detect more singularities of the face image than traditional separable wavelet. The BPCA approach has overcome the low accuracy of PCA in cases of extreme change of facial expression, pose and illumination variations. The face image is first transformed into the wavelet domain using nonseparable wavelet, then the wavelet subbands are divided into sub-images with the same size. The block-based PCA method is then applied to extract features from the sub-images. Finally, weighted Euclidean distance based k-Nearest Neighborhood (kNN) classifier is performed for similarity measurement. Experimental results on the Yale database, the ORL database, and the CMU PIE Database demonstrate that the proposed approach outperforms traditional methods.