Researches on Chaotic Phenomena in Well-logs of Oil-bearing Layers

Finite normal mixture model-based image segmentation techniques can produce robust segmentation results. But the EM algorithm used for learning mixture parameters is very sensitive to initialization. Estimating the number of components and the optimal model parameters inevitably brings a heavy computation burden. In the view of boosting learning, the paper gives a weighted EM algorithm, which is a simple extension to traditional EM algorithm. Using the parameters estimated by the standard EM algorithm as feedback, a resampling-like technique can focus the EM algorithm on those samples which are incorrectly clustered. The experiment results show that the image segmentation results are insensitive to initialization and the number of classes to a certain extent.