Supervised Congealing for Simultaneous Face Normalization and Eye Localization

One of the major characteristics of face images lies in their various facial components (e.g., eyes, nose, mouth) and the spatial regularity among them. Geometric normalization is an important approach to exploit such regularity of face images, by constructing the semantic correspondence between facial features, removing or reducing some variations caused by the changes of pose, scale, expression and so on. In this paper, a novel face normalization method is presented based on the congealing method. Congealing is a recently proposed normalization method which learns a particular affine transformation for each face image such that the entropy of a group of face images is minimized. However, this method does not employ the intrinsic characteristic of face images and needs a time-consuming offline training procedure. We improve on this by training it in a supervised manner and learning the affine transform based on the locations of eyes in a given images, which results in an efficient geometrical face normalization algorithm with eye localizations as byproduct. Experiments on the challenging Labeled Faces in the Wild database show that our method is superior to the original congealing method in terms of recognition performance.