Face recognition committee machines: dynamic vs. static structures

Abstract

We propose a dynamic face recognition committee machine (DFRCM) consisting of five well-known state-of-the-art algorithms in this paper. In previous work, we have developed a static committee machine which outperforms all the individual algorithms in the experiments. However, the weight for each expert in the committee is fixed and cannot be changed once the system is trained. We propose a dynamic architecture on the committee machine which uses the input face image in the gating network to improve the overall performance. In addition, we adopt a feedback mechanism on the committee machine to adjust the weight of an individual algorithm according to the performance of the algorithm. Detailed experimental results of different algorithms and the committee machine are given to demonstrate the effectiveness of the proposed system.