The Research Of Image Segmentation Algorithm Based On CRF

Image is an important medium of information dissemination,in order to use that form of information efficiently,fast and comprehensively,it is necessary for people to do some research on it.Image segmentation is an important research direction of image processing.To understand the semantic information contained in an image,people need to divide the whole image into several meaningful targets firstly.Then,according to each part of the meaningful target people can get the whole image information.How to get the high accuracy image segmentation result has been the hotspot and difficult point in this field all the time.Therefore,the image segmentation algorithm based on conditional random field theory is studied in this thesis.The main contents are as follows:(1)This thesis studied the current image segmentation methods in-depth,from the principle and implementation of various methods on the analysis of the advantages and disadvantages.On this basis,we select the conditional random field segmentation method as the research content.(2)This thesis studied the theory of probability graphs,followed by the study of Markov random field and conditional random field theory,and focused on the parametric learning and model reasoning in conditional random fields.(3)The application of the conditional random field model in image segmentation is realized.The method of conditional random field is used to model the image.The potential function and the interaction potential function are defined in the conditional random field model.Various characteristics of the image that need to be used in defining the correlation potential function are studied.The image segmentation datasets used in this thesis are introduced,and experiments on these datasets are carried out to validate the effectiveness of the conditional random field model for image segmentation.(4)In this thesis,two kinds of higher-order conditional random field models are studied for the problem that the accuracy of segmentation results of pairwise conditional random field model is not high.The higher-order conditional random field model based on non-local region matching defines a higher-order energy function matching the image region marker to constrain the similarity of the region with similar visual features in the image.Another kind of higher-order energy function based on the consistency of super-pixel markup is to incorporate the segmentation quality of super-pixel into the modeling of random field,and add a soft-threshold with super-pixel to measure the segmentation quality.The above two kinds of higher-order conditional random field models have a higher segment accuracy than the pairwise random field model,and its validity is verified by experiments.