Image Classification Based On Probabilistic Graphical Models

Image classification is a very essential task in image processing research. Now, there are varieties of ways to solve this problem. Among them, the probabilistic graphic model based method has been paid more and more attention, and it has provided significant ways to solve the uncertainties in image information field. Recently, the probabilistic graphic model has been widely used in image processing and pattern recognition. In the research of the probabilistic graphic model based image classification, the major problems are the high time complexity of inference and parameter estimation in graph model and the target object in images often can not be extracted completely and accurately from occluded, cluttered and noisy images. The known prior knowledge such as shape prior and the introduction of boundary information are really essential, and they are helpful for directing the target object classification accurately and effectively.This paper focuses on the Conditional Random Field (CRF) and Markov Random Field (MRF) in probabilistic graphic model and their application to image classification. Based on this, three image classification methods have been proposed. The main task of the paper as follows:1. Concerning the high time complexity of inference and parameter estimation in graph model, a superpixel-based conditional random field image classification method is proposed. This method first over segmented the image into small homogeneous regions which were called superpixels using mean shift method. Then the graphical model is constructed with superpixels as nodes. The corresponding definition of CRF and the methods for parameter estimation and labeling inference are proposed and implemented. The experimental results show that better classification results are obtained by the superpixel-based CRF model. At the same time, running time is largely reduces.2. The target object in images often cannot be extracted completely and accurately using only low level image features, such as color, texture, especially from cluttered, occluded and noisy images. A shape prior is introduced into the Markov Random Field model, and a shape-driven MRF classification method for image classification is proposed. The problem of image classification is regarded as finding the maximum a posterior of the labels for image pixels given the image data, which is equivalent to the minimization of a corresponding Gibbs energy function. To exploit the shape prior, the energy function is defined by both image features and the known shape template, and it is minimized by the Graph Cuts method to produce the final labeling results. In addition, an alignment process is introduced to handle the affine variation between the target object and the shape template. The experiments show that the shape plays an irreplaceable role for image classification.3. A kernel principle component analysis shape and boundary based image classification method is proposed. In the method, kernel principle component analysis based shape representation is given, then the registeration problem is sovled and the boundary distance map function is defined. The shape and boundary information are introduced into the energy function, and the whole definition of ernery function is provide. Finally, graph cuts method is used to minimize the whole energy function and the final classification results are given. This method is performed in an iterative way, and the shape prior weight can be adaptively controlled. The experimental results show that kernel principle component analysis based shape prior can capture the shape variations well, and the boundary can restrain the target object as well. Conerning both the shape prior and the boundary of the image to be classified, this method can adept to larger variations between the shape template and the target shape, and receive better classification results.