A PCA mixture model with an efficient model selection method

We propose a new type of principal component analysis (PCA) mixture model which consists of a combination of several PCA models as a good estimation model of complex data distribution. The proposed PCA mixture model has a fast and sub-optimal method of model order selection, so that the time-consuming EM learning procedure is executed only once, with all the PCA bases being kept for a given number of mixture components. Using the ordering property of PCA bases, the effect of PCA bases on a good model is evaluated from an appropriate selection criterion, where each less significant PCA base is pruned, starting from the most insignificant PCA base. As the optimal model order for the given problem, we select a pair of the number of mixture components and the number of PCA bases that results in the smallest classification error over the validation data set. Simulation results of the synthetic data classification and eye detection problem show that the proposed model selection method determines the model order appropriately and improves classification and detection performance