Gaussian mixture models and information entropy for image segmentation using particle swarm optimisation

Wenlong Fu, Mark Johnston, Mengjie Zhang

Abstract: Image segmentation is a key step in image analysis. The Gaussian Mixture Model (GMM) method based on image histograms is popular in image segmentation, but it is difficult to find good parameters of Gaussian models. A hybrid Particle Swarm Optimisation (PSO) algorithm has been used to effectively search the parameters of the models for image segmentation. However, the segmentation results from the fitted models are not stable. In this study, the parameters are optimised by developing a new fitness function in PSO based on information entropy. A combination of the entropy and goodness of the approximation on the image histogram is proposed as a new fitness function further. The results show that the fitness function combining the entropy and goodness of the approximation can be effectively used to obtain better segmentation results than only considering the goodness of the approximation on the image histogram, in terms of intensity errors and the information entropy.