Unsupervised segmentation based on robust estimation and cooccurrence data

The accuracy of any image segmentation method depends on the correct estimation of the parameters of the different regions present in the image, as well as on the correct labelling of the pixels. Using robust estimators and relaxation labelling techniques, an unsupervised segmentation algorithm was developed. The mean gray value of each region is estimated from the histogram using robust clustering analysis. The gray level distribution of each individual region is approximated through the mean gray value cooccurrence data. The standard deviation of the gray levels of each region is estimated from this distribution using the least median of squares (LMedS) robust estimator. The labelling of the pixels is done through an iterative relaxation region growing process, taking into account both spectral and spatial information. The method is tested in various images and validated with synthetic data, where it is shown that the known true parameters are recovered accurately.