Adaptive color spaces based on multivariate Gaussian distributions for color image segmentation

We formulate an adaptive color space for segmenting all image into the two classes "object of interest" and "background" by using well-established methods from statistical pattern recognition. Both classes are modeled by a multivariate Gaussian distribution whose actual parameters are estimated via the Expectation Maximization (EM) algorithm. The output grayscale feature image is derived as the distance of each pixel's color to the decision boundary which is shaped bewteen the two class models. Based on this feature image, which provides a maximum discriminatory power with respect to the underlying model assumptions, the actual segmentation can be performed with appropriate methods from grayscale image processing. This adaptive color space is a practical tool for homogeneously colored scenes, as they appear, e.g., in microscopic images of biotechnical fundamental research.