Image Segmentation With Probabilistic Graphical Model

Image segmentation provides the basis for understanding images,for example,object recognition,image retrieval and so on.However,image segmentation is still a difficult problem,an ideal framework for image segmentation is that it can use a probabilistic method to combine different informations with constraints.This thesis introduces a method for image segmentation,which is based on probabilistic graphical models.The main research contents are as follows:(1)Markov Random Field(MRF)model based on superpixels is used for Image segmentation.First,the SLIC(Simple Linear Iterative Clustering)algorithm is used for over segmentation of the image segmentation,and extracting the superpixels,then MRF model is constructed,which is based on the superpixels and the image observation.And using the Maximum A Posteriori(MAP)method obtains the optimal values of regional nodes.(2)Bayesian Network(BN)model based on over segmentation is used for image segmentation.First,according to the superpixel regions extracts the boundaries,then using the Moravec corner detection algorithm detects the intersections,and according to the boundaries extracts the angles,then BN model is constructed,which is based on the superpixel regions,boundaries,intersections and angles.And using the MAP method obtains the optimal values of boundary nodes.(3)The hybrid probabilistic graphical model is constructed by combining the MRF model with the BN model through the causal relationship between the region nodes and the boundary nodes.(4)The hybrid probabilistic graphical model is transformed into Factor Graph(FG)by factor graph theory,and the optimal value of the random variables in the image are obtained by the max-product algorithm,and for Image segmentation.MRF model only captures the spatial correlation between random variables,BN model only captures the causal dependencies between random variables.However,the hybrid probabilistic graphical model captures the spatial correlation and the causal relationship between the variables at the same time.The experimental results show that the hybrid probabilistic graphical model improves the accuracy of image segmentation.