Segmentation Of Color Images By Unsupervised Bayes

MRF(Markov Random Field) combined local information and spatial information has been widely used in machine vision and image process related area. However MRF is NP Hard. As MRF and Gibbs distribution of equivalence being put forward, then computing has become possible. In this paper ,we use MRF to segment color image. As the original MRF is supervised, we introduce a novel approach called Dirichlet Markov Random Field for color image segmentation. The approach uses Dirchilet Process Mixture (DPM) to get local information of MRF's energy function. Instead of finite component, DPM use infinite component, in which the prior distribution is defined on the space of all possible distribution.In this paper,the Dirichlet Process Clustering algorithm performs Bayesian mixture model. The idea is that we use a probabilistic mixture of a number of models that we use to explain some observed data. Each observed data point is assumed to have come from one of the models in the mixture, but we don't know which one it is. The way we deal with that is to use a so-called latent parameter which specifies which model each data point came from. Yet how to assure the number of models becomes the core of the problem. The probability model will not state complex distribution if models are chosen too little. And in contrary, if chosen too many, it will cause over-fit problem. So we choose Dirichlet distribution to choose the prior probability. Dirchlet distribution is the one based on multi-parameter, and can be regarded as generalized forms of beta distribution. Derichlet distribution is the distribution of distribution, which means that a sequence of samples from the Dirichlet distribution is a discrete distribution space. Given k as the number of models, DP still stands when kâ†'∞. Because the number of K is related to the number of observational data, so it belongs to Nonparametric estimation. Also the number of K Decreases exponentially, so generally it will not rise over-fit problem. In optimization problem, the issue uses Variational inference of blei's paper to solve DP mixed model problem, through which, we can deal with the problem of clustering images in acceptable time. How to measure two distance function of pixel is an important step to decide accuracy that based on color segmentation. The original method is to use Euler distance. However, this paper consider the method not suitable for visual habits. Through observation and experiment to color space, we find that relative to the Euler distance, the angle distance between the two pixels plays a more important role. However, the distance itself exits faults. Such as pure black and pure white pixel, if using angle distance to compute, it will be considered the two pixels completely similar. In order to overcome the above problems, the paper relies Euler distance and angle distance, also Combines the characteristics of them to design a distribution based on them. The paper makes use of the distribution for Dirichlet Process as distribution model. For the reason that this issue takes consideration of more appropriate comparison of model pixels and local information space, thus gets better segmentation.In order to solve the Large volumes of data in image , This paper explores parallelism of Yuri Boykov's Graph Cut algorithm and designs a parallel algorithm for Grow and Adopt function. The branch and bound method is used for Breadth First Search in Grow function and extend the cancel thread to improve the run times. The new feature'task'in OpenMP3.0 is used to solve the irregular problem. For p CPUs, the worst case reduce the running time from O(bd/2 +a+nr) to O(bd/2 /p +a/p+nr/p). Implement the algorithm using OpenMP on 1 and N CPUs .The experiment proved this algorithm is precision and efficiency.