USING COMPETITIVE PROTOTYPES FOR THE VERIFICATION ON UNSPECIFIC PERSON

摘要

For face verification on unspecific persons, most approaches are based on the single similarity between the test face and the claimed client face. This paper proposes a novel verification algorithm using competitive prototypes, including both the claimed client face and the competitive negative samples representative of other classes outside the client set. The test sample is matched with each competitive prototype, and matching scores are combined for a final decision. Three schemes of applying competitive prototypes are designed, tested and compared with the similarity-based verification approach on several databases, features and classifiers. Experiments demonstrate that they are useful for the verification on unspecific person and they improve the verification performance by reducing the error rate.