A study of supervised, semi-supervised and unsupervised multiscale Bayesian image segmentation

In this paper, we study multiscale Bayesian image segmentation with respect to the different availability of image features. Specifically, wavelet domain hidden Markov models (HMMs) are adopted to obtain statistical image characterization. The joint multi-context and multi-scale (JMCMS) approach is also applied to exploit robust multiscale contextual information. We first review the supervised Bayesian segmentation algorithms where complete image features are given. Secondly, we study semi-supervised segmentation by only providing partial image features. The K-mean clustering is used to convert the semi-supervised segmentation problem into the self-supervised process by identifying the reliable training samples. Thirdly, an unsupervised segmentation algorithm is also developed where image features are completely unknown and can be trained online during segmentation. The simulation results on a synthetic mosaic show that the proposed algorithms can achieve high classification accuracy for both semi-supervised and unsupervised segmentations.