Image Segmentation Based on Spatially Coherent Gaussian Mixture Model

It has been demonstrated that a finite mixture model (FMM) with Gaussian distribution is a powerful tool in modeling probability density function of image data, with wide applications in computer vision and image analysis. We propose a simple-yet-effective way to enhance robustness of finite mixture models (FMM) by incorporating local spatial constraints. It is natural to make an assumption that the label of an image pixel is influenced by that of its neighboring pixels. We use mean template to represent local spatial constraints. Our algorithm is better than other mixture models based on Markov random fields (MRF) as our method avoids inferring the posterior field distribution and choosing the temperature parameter. We use the expectation maximization (EM) algorithm to optimize all the model parameters. Besides, the proposed algorithm is fully free of empirically adjusted hyperparameters. The idea used in our method can also be adopted to other mixture models. Several experiments on synthetic and real-world images have been conducted to demonstrate effectiveness, efficiency and robustness of the proposed method.