The Research And Application Of Conditional Random Field

The task of image segmentation is dividing images into several non-overlapping and meaningful areas which have same of similar natures, such as similar texture or similar gray levels. Image segmentation is the basic premise of image recognition and image understanding, because the quality of segmentation is directly affects subsequent image processing, so image segmentation is very important.As a kind of collection for mode, Probability Graphical Model using graphic mode to express modes based on the probability correlation relationship, which is a graphic structure based on Bayes rules in the theory of probability, at the same time which is also the most popular graphic model in the area of artificial intelligence. The model is first put forward by Whittaker based on some theory in statistical theory, which can be used for predicting by collecting context relationships and has a good performance. Such model provides a solution for solving uncertain problem in artificial intelligence. Because PGM put the information of time and space together in actual problem, and based on a large number of independent relationships between variables so that it can construct a structure model to effectively express data information based on joint probability distribution. Probability Graphical Model is divided into directed graph model and undirected graph model, while Conditional Random Field is a undirected graph model. CRF is first used for natural language processing, and used for image processing later, such as denoising, segmentation etc.In the view of the above, this paper using CRF to segment image in two image databases, a serial images from Brainweb to segment white matter, gray matter and cerebrospinal fluid, Weizmann horse to segment horses from a variety of color images. The main research results of this paper are as follows:(1) We proposed a segmentation method, which using superpixel and modified conditional random field, regarded as S-MCRF, we also discuss and research the method preliminarily, and this method provides a certain reference for image segmentation in Probability Graphic Model’s practical application.(2)When we segment MR image into superpixels, some edges of superpixels can’t be getting close to the edge of different tissues, such as the narrow area of white matter and the cerebrospinal fluid among gray matter. So in order to improve the precision of segmentation, we need to use pixel instead of superpixel, however this will lead to a great improve of the time complexity, because the number of nodes in graph decides the time complexity of algorithm, so we need to select using superpixel or pixel according our needs.(3)As for the images in Weizmann horse database, we’d like to use superpixel instead of pixel, because the target horse in different complex environments, using superpixel as nodes has more advantages, such as lower time complexity, and more integrity.