Robust gaussian mixture modelling based on spatially constraints for image segmentation

Gaussian mixture model (GMM) has been successfully applied to image seg- mentation. However, the drawback of GMM is that it lacks of robustness against noise for image segmentation. To effectively reduce negative effects of the noise, in this paper, we propose a variant of GMM which fully considers the spatial relationship between the pixels and the label probability proportions are explicitly modelled as probability vectors. At the same time, the component function of a pixel is also closely relative to its neigh- boring pixel. In the inference process, gradient descend method is adopted to estimate the parameters of the proposed model. The proposed model compares with some models which are related to mixture models. Several experiments are conducted on both synthetic grayscale images and real-world natural images. The experimental results show the ro- bustness and accuracy of the proposed model outperform some state-of-the-art models. © 2015 Ubiquitous International. All rights reserved.