Simple Image Segmentation using Gaussian Mixture Model
- Extract the feature vector (e.g. Color) for each pixel in the image;
- A clustering method is applied in the feature space to cluster the data into several clusters each of which corresponds to one homogenerous region in image;
- According to the cluster centroid, each feature vector is assigned to a cluster and the same as the corresponding pixel;
Two type of classification methods can be used: Non-parameteric Method and Parametric Method. Gaussian Mixture Model (GMM) is a parametric clustering method. It assume the distribution of data is mixture model of several distributions whose forms are known (gaussian). In other words, the distribution of the whole data is a linear combination of several gaussian distribution (mixture coefficients). A Expectation Maximum (EM) algorithm is used to estimate the unknown parameters. Firstly, the data are randamly assigned to these gaussian classes, and the mean and variance of each gaussian are calculated in current iteration (E step). According to the probability of different classes, the data is reassigned (M step). This procedure iterates until convergence and the parameters are estimated. Using the probability mess function (pmf), the data can be classified and thus pixels are segmented into several regions. The following images show an example.
 |  |
| Original Image | Segmented Image |
