Automatic keyface selection for known people identification in images

Abstract

We propose a set of features to characterize faces in images. The goal is to use these features to automatically select the most relevant images to train an identification tool. Those features are derived from a set of constraints usually required to allow the recognition process. A filtering tool based on the Adaboost algorithm is used as a basic process to test the relevance of these features for such a task. In these experiments we obtained a rate of 87% of good selection. In other words, among all the faces kept after the filtering process, 87% are compliant with the predefined constraints, and can be used to train an identification tool.