Two-step supervised confidence measure for automatic face recognition

This paper deals with automatic face recognition in the context of a real application for the Czech News Agency. This system will be used to annotate people in photographs during insertion into the database. Unfortunately, the accuracy of the current face recognition approaches is limited and therefore another task to process the recognition results is very important. The main contribution of this work thus consists in proposing and evaluating a novel supervised confidence measure method as the post-processing step in order to detect incorrectly classified face images from the classifier's output. We experimentally show that the proposed confidence measure is beneficial for our application.