A supervised probabilistic model for visual object recognition

Abstract

In this paper, we propose a new supervised probabilistic-based method for object recognition. Specifically, we conduct a process of supervised learning in which each class is represented by Gaussian Mixture Model (GMM). In order to group images having the same visual appearance, we cluster images related to each class using k- means. Probability density functions that correspond to the resulting clusters are fused in a GMM representing the class model, where Expectation-Maximization (EM) is used to estimate the parameters of the mixture. Given a test image, we calculate the probability that the image belongs to each class. The image is then assigned to the class having gained the highest score. The proposed method takes into account the intra-class variation and it is capable to distinguish different objects in spite of the small inter-class variation. Experimental results demonstrated the effectiveness of our method and an accuracy of 95.28% is reached.