Robust Image Segmentation Via Bayesian Type Criterion

Information criterion is a traditional method in variable selection problem as it can choose the best variables from all the candidates and simplify the model, further analysis will be easier by using it. While, even if its theory and conclusion is perfect, traditional information criterion is very strict with its error, and when the error is nonrandom or unknown in practice, it may not get good result by using these methods directly. To deal with the nonrandom or unknown error, a robust Bayesian information criterion (RBIC) is proposed based on the traditional information criterion. Image segmentation is one of the most important steps of image process in computer science, and objective image often contains different errors in practice, thus may influence the accuracy of result. We apply our robust information criterion into image segmentation method to make sure the segmentation number and its segmentation regions, handle the negative factors of different error. The asymptotic properties are also studied with some conditions. Moreover, Monte Carlo numerical experiments are conducted to evaluate the performance of the new method comparing other information criteria by combining our information criterion with a designed image segmentation method.The rest contents of the paper is as follows. Section1is the introduction of information and image segmentation, the existing conclusion of image segmentation and robust estimation method are also included. Section2describes our robust Bayesian information criterion, includ-ing the assumed condition and the asymptotic properties of our model. Section3provides a new method of image segmentation, two Monte Carlo simulations and a real image simulation are also studied. The paper ends with a brief discussion.