Unsupervised Image Sequence Segmentation Based on Hidden Markov Tree Model

This paper presents a novel unsupervised image sequence segmentation method using hierarchical wavelet domain hidden Markov tree model(WDHMT). The key idea is that with a priori information introduced into the segmentation framework, we can capture both local and global statistical information using WDHMT model. Firstly, each frame extracted from the image sequence is transformed through discrete wavelet transform(DWT) to obtain a compressive representation of the original one. Then we capture the context information of wavelet coefficients at each level through tree-structured probabilistic graph. After the model parameters are learned through up-down iterated expectation maximization(EM) algorithm, we-deduced the maximum likelihood(ML) segmentation at the finest level. The boundary information is then fused with the a priori region information. Finally, we quantitatively evaluated the performance of this algorithm by using a sequence of tobacco leaf images polluted by Gaussian white noise. The simulation results show that the proposed algorithm can achieve high classification accuracy, preferable specificity and sensitivity properties.