Face Verification Based on Feature Transfer via PCA-SVM Framework

Abstract

Face verification is a task to determine whether a pair of given facial images belong to the same person. In unconstrained real applications, inter and intra variations, including illumination, pose, occlusion, and expression, will seriously decrease the verification performance. Due to the lack of annotated data for face verification, extended datasets for face recognition with large samples are used to assist learning a robust feature representation generally. However, the extended data for face recognition is different from face verification on distribution and task. In this paper, a transfer learning based on PCA-SVM is proposed to alleviate above problem. The original feature representation is learnt from a deep convolutional neural network by face classification. Then a PCA-SVM based transfer method is used for feature reprojection from the source domain (face recognition) to the target domain (face verification), which reduces the divergence of feature distribution and task inconsistency. The proposed framework yields comparable results and the accuracy is 98.5% on LFW dataset.