有约束的自适应联机手绘图形识别技术研究

On-line hand-drawn graphics technology is important in intelligent human-machine interface to improve intelligence of human-machine interface. In this project, theories and methods of adaptive recognition of on-line hand-drawn graphics under constrained circumstance are researched. In the proposed recognition system, an on-line input graphic symbol is recognized by a two-layer primitive-symbol structure. Through the recognition of pre-defined primitives, the searching space of recognition algorithm is greatly reduced. The set of primitives is selected based on the knowledge of graphics symbol, so the types of primitives are constrained. Where, primitive types are decided by the apriori knowledge of graphics set and are constrained..In the phase of primitive recognition, an adaptive HMM (AHMM) recognition structure with feedback is presented. The primitives can be recognized adaptively through a mechanism, which combines adaptive feature compression algorithm and a closed loop feedback technique. In order to improve the recognition speed and accuracy, a bi-layer multi-resolution recognition structure based on AHMM (BL-AHMM) is also proposed, the genetic algorithm is also introduced into this recognition system for HMM training. The BL-AHMM network is proved to be time efficient and can achieve much higher recognition accuracy than normal HMM recognition system..In the recognition of complex graphics based on primitives, an error tolerant matching method based on the order constrained attribute relation graph (OCARG) is proposed. This method use attribute relation graph (ARG) to describe the relation between primitives and graphic symbols are designed. Under the constrained of primitives types, the constrained attribute relation graph is constructed using vertexes sorting strategy to describe relations between input graphics and standard graphics. A fast matching algorithm for OCARG is presented. The similarity measure between vertexes and edges of CARG are defined to realize the fast error tolerant matching and improve the intelligence of the whole recognition system and make the human-machine interface friendly..In this project, theories and designing methods of optimal classifier is researched also. A novel concept of Voronoi Diagrams, Multi-color Voronoi Diagrams, is proposed. Several important features of Multi-color Voronoi Diagrams are given also. Multi-color Voronoi Diagrams is the coloration result of Voronoi Diagram based on data set characteristics. A classifier with gradually local learning characteristics based on Multi-color Voronoi Diagrams, Multi-color Voronoi classifier (MCVC), is proposed. MCVC is an unlinear optimal classifier and can be used to linear and nonlinear classify problems. In the meantime, it has ability to learn locally from new data samples rapidly, which overcomes overfiiting problem in neural network method. This ability is similar to the ability of human being. Multi-color Diagrams can be used to pre-select support vector in the construction of support vector machine, A bi-color Voronoi to pre-select support vector is proposed also, which can improve learning speed of support vector machine..

有约束的自适应联机手绘图形识别技术着重研究在待识图形集和基本笔划受限的情况下，对用户输入单笔划时的输入习惯的自适应学习技术，对子图识别中基本线条的类型和属性的自适应调节能力，对用户输入的图形中出现频率较高的新符号的自适应学习能力，不完备信息的模糊约束属性关系图匹配技术。本项目对提高人机交互的智能化水平有很重要的意义。