Key Technologies In Example-Based Pattern Generation

Patterns are ubiquitous in our world and reflect human being's pursuit to beautiful and happy life. With the development of society and improvement of technology, the methods of pattern generation are changing from manual mode to a mode that makes use of computer in design. Currently, there are three kinds of method of using computer in pattern design:CAD, mathematical-based and knowledge-based design. They improve design efficiency, but have their own limits in some applications. For Instance, It is difficult to construct accurate mathematical models or rules description for floral patterns that are free and variable in form.Based on surveys of related work on pattern generation, a new pattern generation method that makes use of example learning and feature matching techniques and a noval object feature describtion model are presented to deal with the difficulties lies in floral pattern generation. This thesis also proposes to introduce human visual experiences similarity measure by means of training data to improve the accuracy of similarity matching. The main contributions of the thesis include:(1) As far as sketch and primitive motif who only have edge information are concerned, a noval contour based ray shape model are proposed, it represents object with local features and spatial relations of sample points on contour, a maximum voting scheme is used for object matching and obtaining geometrical transforming invariance. Experiments demonstrate the robustness in representing shape feature of object, and can be used for both contour matching and object detection in edge map of image. Then, an unicity-based feature selection algorithm is designed to find a more unique feature to represent flower image, and a NN-based (Nearest Neighbour-based) image object matching method is proposed. With the optimal feature descriptor and matching algorithm, experiments show good results on floral image object matching.(2) Similarity measure ia basic problem in pattern recognition and computer vision, and classic methods that measure object similarity using distance between two objects. But distance only conveys geometrical similarity instead of similarity perceived by human. A human vision related factor(HVRF) based on the statistical characteristics is proposed, then human visual experiences are introduced into similarity measure by HVRF to improve the accuracy of matching. Considering that vision is fuzzy, we emphasized on using membership degree as HVRF, and present feature-space-division-based and fature-value-distribution-based methods of constructing membership function. Experiments show that similarity measure with HVRF is more consistant with human perception.(3) Art style is difficult to be described accurately. To solve this problem, an example-based pattern generation method is proposed. This method doesn't need extract domain knowledge model, the learning process is performed by collecting example patterns, and execution process is matching input object to examples. Two pattern generation mode and their corresponding workflows are presented. On one hand, pattern can be generated from manual sketch that directly reflect designer's inspiration and intention, on the other hand, we provide user with an easy way of translating floral image to stylized pattern. The method decreases the complexity in pattern design and helps user design pattern easily.