Feature selection in interactive face retrieval

In this paper, we introduce a novel perspective to feature selection in face retrieval, for the purpose of increasing the coherence in face similarity feedback, and hence narrowing the semantic gap between human and machine in face perception, and thus speeding up the interactive retrieval. The coherence is defined in previously established interactive face retrieval models and the feedback database of human users is built up for building or testing those models. Based on this coherence database, we propose a novel criterion function as the target function in feature selection to measure the coherence. Four feature selection algorithms have been adopted in comparison: the minimal- Redundancy-Maximal-Relevance criterion (mRMR), the best individual (BI), the sequential forward selection (SFS), and the Plus-l-Minus-r (l-r). The l-r algorithm proves to be better than the other three methods in feature selection for coherence. Experiments demonstrate that the proposed feature selection method could largely improve the interactive searching efficiency and face recognition rate in face databases.