Study On Automatic Portrait Generation Based On And-or Graph Representation

Automatic human portrait generation is one of challenging problems in computer vision community. Human portrait includes many patterns, while each pattern contains complex structure and configuration. So it is difficult to find a universal model fitting to all possible patterns. This paper presents an automatic human portrait generation method based on the And-Or graph representation. Our research focuses on the image understanding and template matching.An improved Active Shape Model (ASM) algorithm is proposed to extract facial feature. Facial feature is important geometry information for portrait generation According to the disadvantage of ASM, we make progress in its local texture model, search algorithm, and model initialization. Experimental results show the improved algorithm is more efficient and accurate.Next, an automatic human portrait generation method of based on the And-Or graph representation is presented. The And-Or graph is a generative model, which separates the structure and style of portraits and accounts for the variability of portraits. The method decomposes portrait into three components: hair, face, and collar. Each component has a number of distinct templates. Given a face image, the best matched template of each component is selected by template matching. Guided by the And-Or graph for portrait, a new portrait made of a set of matched templates is generated. The method benefits from a number of template dictionaries for portrait components in different styles. Therefore, it is convenient to change the styles of portrait by changing the template dictionary. A number of experimental results demonstrate the effectiveness of the method.Further, we present a method of facial components classification by integrating subspaces selection and clustering. The proposed method contains two iterative steps: (1) Fixing the subspaces, k-means clustering is performed to generate cluster labels; (2) Fixing cluster labels, the samples are projected onto lower dimensional subspaces to do subspaces selection through Linear Discriminant Analysis. Through iterative steps, clusters are discovered in the lower dimensional subspaces to avoid the curse of dimensionality, while the subspaces are adaptively re-adjusted for global optimality. We also apply the classification results to the facial components matching of portrait generation.Moreover, we present a novel method for hair classification, matching and sketching. We extract shape and appearance features and calculate statistical properties from the training data. Base on these features, we learn twenty-four hairstyles. For a given hair images, the best matched template is found by using a cascade pruning algorithm. Taking the template as a prototype, a new hair sketching corresponding to the given image can be generated by a curve fitting and texture mapping algorithm. We also present a template matching method for collar based on Shape Contexts. We achieve efficient and accurate matching results by comparing the Shape Contexts distance between the collar template and the primal sketch of collar image.At last, we summarize the presented work. According to the imperfect aspects, we analyze and discuss the future work.