Smoothness Assisted Interactive Face Annotation via Neural Network.

Face annotation aiming to tag faces with identities, is an essential tool for image retrieval and management of character-centered photo albums. Conventional face annotation systems have a demand for fully labeled training data, which is hard to get in that manual annotation is a tedious and high-cost work. The aim of our work is to reduce as much as possible manual workload in face annotation. Toward that end, we proposed a smoothness-based model for semi-automatic face annotation, which first applies smoothness constraint and context information such as co-occurrence relationship(e. g. two faces extracted from one photo must have different identities) to a neural network and then determines a candidate list of faces that need to be annotated manually on the basis of active learning strategies. Experimental evaluations on two real photo albums show the effectiveness of our proposed model.