An Unsupervised Learning Algorithm for Image Segmentation Based On Finite Mixture Models
Linsen Yu, Zhang Tianwen · 2006
There are two open problems for unsupervised learning of finite mixture models: model selection and initialization. To circumvent these problems in image segmentation applications, we integrate a filter technique into the EM algorithm. The proposed algorithm starts with the largest possible number of image regions. With the convergence of the algorithm, irrelevant components can be eliminated. It does not require careful initialization and also has the advantage of preserving the good features of EM while making use of the spatial information in a reasonable amount of time.