The work in this paper is supported by the following projects :

1 es in the following wo er, BN classifier is the most suitable for multi-stroke symbols and per Sketch is a type of natural, simple and convenient human-computer interaction which conveys ambiguous conceptions and imperfect information. It adapts to creative activities and improves the collaboration and communication among people. With the rapidly extended application of hand-written characters and hand-drawn sketches, online sketchy symbol recognition has been widely studied, in the process of which symbol feature representation, intelligentized recognition approaches, learning ability and the extension to three-dimension space have been the key points of this area. Regarding these issues, this paper engag rk: Firstly, regarding the lack of adaptation to the users’ drawing styles of recognition approaches in sketchy symbol recognition, this paper proposes intelligentized recognition methods. We choose three typical machine learning methods, including support vector machines(SVM), hidden Markov models(HMM) and Bayesian belief network(BN), and then design the sketchy symbol-based models and the corresponding recognition strategies. We compare their accuracies and efficiency, analyze the relationships among recognition methods, feature representations and users’ drawing styles, from the aspects of multi-users’ samples and different sizes of samples respectively. The experimental results show that symbol features are similar in representing abilities for different users’ drawing habits and greatly influence the accuracies and training time of recognition methods. Moreov forms best in training time. Secondly, regarding the problem of “collecting sufficient samples and maintaining their distinction” in the intelligentized recognition methods, this paper introduces artificial immune theory to sketchy symbol recognition and proposes innovatively a clonal selection algorithm based on detector generation to solve the problem of limited sam le learning in sketchy symbol recognition. This algorithm generates initial detectors with a p − r contiguous-bits unchanged rule and − p receptor editing to search in a wider feature space and try to avoid local convergences. We consider hand-written Chinese characters as the experimental samples, and analyze the influence of the training parameters. The experiments show the improvements of the training process and the classification results of sketchy symbol recognition. Thirdly, regarding the problem of sketchy symbol based three-dimensional(3D) shape representation, this paper generally analyzes and compares the 3D shape descriptors commonly used in 3D object retrieval and employs several typical ones to 3D shape recognition. The experimental results show that as for the 3D shape descriptors with well descriptive abilities, all of their combined descriptors perform better than each single shape descriptors in descriptive bilities and recognition accuracies.