An Improved Greedy Search Algorithm of Bayesian Network Structures for Facial Action Units Recognition

Abstract

Bayesian Networks (BN) are an effective method to recognize facial action units (AUs) combinations, which is a key issue of AUs recognition. Learning BN structures from data is NP-hard. Greedy search algorithm is a practical approach to learn BN from data, but it is liable to get stuck at a local maximum. In this paper, an improved greedy search algorithm is proposed in order to deal with the above-mentioned problem. The proposed algorithm starts from a prior structure, which is constructed by prior knowledge and simply statistics of AUs database, then updates the prior BN structure not only with the BN structure that has maximum score among all of the nearest neighbors of the prior BN structure, but also updates it with some BN structures that have higher score. The experiments show that the proposed algorithm is computationally simple, easy to implement, and may effectively avoid getting stuck at a local maximum.