Texture analysis is an important research area in computer vision, image processing and pattern recognition. In order to deal with texture problems, we usually make a texture model and extract features. In this thesis, after a brief review of the related work, a Bayesian network texture model is given based on which texture synthesis and classification methods are proposed. The content of this thesis includes texture synthesis based on Bayesian networks, Bayesian network texture classification and shape from texture using scale invariant features. Firstly, a texture synthesis method based on Bayesian network model is given. Texture synthesis is an active and useful research area, it can be used to check whether a texture model is proper. Since Bayesian network is good at describing the causal and dependency relations between nodes, in this thesis, we use it to model the relations between texture pixels in a small region and propose synthesis methods for binary, gray level and color textures. Secondly, a Bayesian network texture classification method is proposed. After modeling a texture image with Bayesian network, features can be extracted from the model. Pixel values in a small image patch are treated as a realization of a Bayesian network. All the patches form a sample from which the network structure and parameters for conditional probability distribution functions (CPDs) of nodes are estimated. Then the histograms of these CPDs for each node are used as features for texture classification. Finally, shape from texture problem is solved using scale invariant features. The direction of a texture plane can be estimated by the scale variance of the alike local structure patterns. After identifying all the key points which have scale invariant features, they are clustered by their SIFT descriptors. Under the assumption that all the local structure patterns in the same cluster have the same size, the direction of texture plane is estimated by the change of their scale. Robust IRLS method is used to make up with the error of clustering process.
修改评论