Pose-invariant features and personalized correspondence learning for face recognition

In surveillance systems, face recognition plays an important role for human identification. In such systems, human faces are spatially unconstrained, which results in a significant change in pose, and face recognition becomes more challenging when only one frontal image of the face has been registered in the gallery. In this study, we attempt to solve the problem where only one frontal image of the face is registered in the gallery, and the probe faces are captured in unconstrained poses. The face likelihood is measured using pose-invariant features of scale-invariant feature transform (SIFT) and personalized correspondence learning method. A generic correspondence is first learned between the poses, and the pose-invariant SIFT is fulfilled by extracting the keypoints on virtual patches that are generated by a generic correspondence with the pose variation. The generic correspondence is further personalized to fit each subject, and the learning error of the personalized correspondence is combined with pose-invariant SIFT to measure the face likelihood. The experimental results indicated that our proposed algorithm achieved an average performance of 95% across poses within 40 degrees, which is better than other well-known algorithms.