Towards a better comprehension of the role of image registration in face recognition algorithms

Abstract

Automatic face recognition systems have currently reached high hit rates. Nevertheless, simple steps like image registration are not being considered in several methods. The alignment of the set of images in a same coordinated system must be seen as an initial and crucial step in algorithms that are based on dimensionality reduction. This work aims at analyzing the importance of registration as a preprocessing step in recognition algorithms based on eigen decomposition. A set of experiments was conducted, in which images of publicly available databases were processed under rotation and translation. These operations put the images on controlled non-registered form for precise evaluation of three recognition methods — Principal Component Analysis, Two-dimensional Principal Component Analysis and FisherFaces, combined with the Euclidean distance and cosine similarity measurements. The experiments revealed the best combination of methods. Moreover, the behavior of hit rate with respect to rotation and translation misalignments were characterized as a Gaussian function.