The Nonparametric Method Based On The Theory Of The Mutual Information Analysis And Its Application In Image Segmentation

Image segmentation is a kind of image processing technology that it divides the image into different areas with various characteristics and extracts the interested part from the image. It is the foundation of image visual analysis and pattern recognition. Moreover, it is also a classic and difficult problem. So far, there is not an general image segmentation algorithm suitable for all images. Therefore, the purpose of designing more generic, efficiency and accuracy algorithm become more popular for some scientists.Nonparametric method based on the mutual information theory is a new method of image segmentation. In this method, the image is divided into the internal and external region label by the curve, so we can cast the segmentation problem as the maximization of the mutual information between the region label and the image pixel intensities, subject to a constraint on the total length of the region boundaries. We solve the information-theoretic optimization problem by deriving the associated gradient flows and applying curve evolution techniques. We use the level-set method to implement the resulting evolution.The experimental results based on both synthetic and real images demonstrate that the proposed technique can solve a variety of challenging image segmentation problems. Because the original algorithm efficiency runs too low, we use narrowband level set method and the Fast Gauss Transform to improve the original algorithm, the improved algorithm keeps the advantage of the nonparametric algorithm and has a high computing efficiency.First, we summarized the traditional image segmentation method and its shortcomings, presented the basic theories of the non-parameters algorithm, and introduced domestic and international research status in the image segmentation field.Second, we studied the nonparametric image segmentation method based on the mutual information. This method described the energy function of image segmentation with information theory. We get the energy function of image segmentation by estimating the maximization of the mutual information between the region label and the image pixel value. One of the advantage for the method is that it does not need to set the parameters, another is more kinds of images can be dealt with, furthermore, this method is not sensitive to the noise in the image.Last, we improved nonparametric method with the narrowband level set method and the Fast Gauss Transform. Nonparametric method is an iterative calculation, in the process of minimizing the energy function, we should estimate the mutual information between the whole region labels and all the image pixel values in one iteration. And this lead to a large amount of calculation. Even we use narrowband level set method, which makes the level set function update only in the narrowband, reduce the computation time, the efficiency of the algorithm is still very slowly. Then we introduce the Fast Gauss Transform to reduce the time used for estimating the mutual information. Compared with the traditional algorithm, the improved algorithm keeps the advantage of the nonparametric algorithm, which does not have parameter selection problem and it is also not sensitive to noise in the image. The algorithm efficiency is greatly improved after the improvement, this makes the nonparametric method be more suitable to practical problems.