What is a Proper Face Registration for Face Recognition?

Face recognition systems have now made significant progress. In practice, we may face the important question of what is a better face registration for a typical face recognition algorithm, but this problem has often been overlooked in the past. In this paper, we aim to answer this question by investigating the relationship between different face registration methods and the performance of different face recognition algorithms. Both the global and the local affine transform based face registration methods are investigated. Our conclusions are that a more accurate face registration is not necessarily for different face recognition algorithms. In most cases, a face registration based on the global rigid transformation with three landmarks works well enough. On the contrary, an excessive registration based on the non-rigid transformation with a large number of landmarks can lead to a distortion of the face structure and then weaken the recognition performance. Furthermore, the face contour is important for distinguishing different faces and should be preserved in the cropping stage of face registration.