Learning of Finite Two-Dimensional Beta Mixture Models

Finite mixture models are widely applied in various domains of applications. They assist to analyze datasets, achieve better insight into the nature of data, discover latent patterns and provide critical knowledge that we are looking for. The main focus in past works was on Gaussian statistical models, however there are several applications involving asymmetric and non-Gaussian data. Parameter estimation is one of the essential and fundamental challenges of statistical researches. Some deterministic approaches such as expectation maximization (EM) have been mainly considered as effective techniques to deal with this issue. In this article, we introduce a bivariate Beta distribution with three parameters as our main parent distribution which could be applied in skin detection and image segmentation. The feasibility and effectiveness of the proposed method are demonstrated by experimental results that concern both artificial and real datasets.