Face Verification Using High Dimensional Feature and Metric Learning

Abstract

Face analysis (face detection, recognition and facial expression recognition) has a large number of applications, including security check, faces beautification, customer attribute analysis and computer entertainment. Although research in face recognition has been conducted since the 1970s, this problem is still largely unsolved, and the recognition performance gap with human is still large. Currently, among all the face analysis techniques, face detection already becomes relatively mature

the remaining issue mainly lies in robust face recognition and facial expression recognition. Even recent years have seen significant progress face recognition algorithm, but reliable face recognition still offers a great challenge to computer vision and pattern recognition researchers, especially in the unconstrained environment with no uniform control with pose, illumination and facial expression. Along with the release of Labeled Face in the Wild (LFW) database, recently there is extensive increased interest in face recognition. With this large face database captured in the wild, there are some new research directions in face recognition, for example, developing face verification algorithm instead of face identification algorithm, and numerous researchers have reported their accuracy based on LFW. Within the last two years, the performance on LFW has been boosted significantly, made it possible to develop robust automatic face recognition system which is comparable with human ability. In this thesis, we propose a novel face verification algorithm and obtain the performance which is comparable with the current state of art performance to some extent. Our approach utilizes a powerful local image descriptor, i.e., high dimensional ULBP (Uniform Local Binary Pattern), which can capture enough details of image without increasing too much computation cost on feature extraction. Based on this descriptor, we combine the transfer learning based Joint Bayesian model and metric learning approach to learn the final classifier. In the transfer learning based joint Bayesian model, we first use a large external database to learn the source joint Bayesian model, and then transfer this knowledge to the target database. To improve the discriminative power further, we optimize the matrix obtained from the transfer learning based joint Bayesian model with the goal that larger the gap between the distance of the intra and extra face image pairs, to make the optimization tractable, we alternatively optimize the matrices in an EM fashion. The evaluation result based on LFW dataset showed that the accuracy of our algorithm is comparable with the current state of the art, and it’s only slightly lower than several recently released commercial vendors, for example, algorithms developed in Facebook and Face++, both of which deploy the deep neural network and use extremely large external training dataset. Moreover, our algorithm can be further optimized to reduce the computation time in test time.