Automatic Images Classification Using HDP-GMM and Local Image Features

摘要

In this paper, we propose a new method based on the probability model that can classify automatically various images by subjects without any prior exchange of information with users. First, we introduce the hierarchical Dirichlet processes Gaussian mixture model (HDP-GMM) that can be applied in images classification, and consider the variational Bayesian inference method to estimate the posterior distribution for the hidden variables and parameters required by this model. Second, we examine the extraction method of various local patches features from given image, which can accurately represent the colors and contents of images. Next, we have trained the HDP-GMM using the extracted patch features, and then present a scheme to classify a given image into the appropriate category or topic by using trained model. Finally, we have applied our model to classify various images datasets, and we have showed the superiority of the proposed method using several evaluation measures for classification method.