Optimal segmentation of signals and its application to image denoising and boundary feature extraction

An optimal procedure for segmenting one-dimensional signals whose parameters are unknown and change at unknown times is presented. The method is maximum likelihood segmentation, which is computed using dynamic programming. In this procedure, the number of segments of the signal need not be known a priori but is automatically chosen by the Minimum Description Length rule. The signal is modeled as unknown DC levels and unknown jump instants with an example chosen to illustrate the procedure. This procedure is applied to image denoising and boundary feature extraction. Because the proposed method uses the global information of the whole image, the results are more robust and reasonable than those obtained through classical procedures which only consider local information. The possible directions for improvement are discussed in the conclusion.