Study Of Recognition For Hand-drawn Elec-tronic Component Symbnol Based On Radia Ba-sis Function Neural Networks

Due to the advantage of simple, fast and conveniently storage, CAD technology plays an irreplaceable role in various fields designing. It can greatly enhance the qual-ity of design, reduce the design cycle, share device resources and strengthen data han-dling capacity. But CAD technology is mainly applied in early stages of design and has no role in catching inspiration and thinking exploration. It can't meet the needs of early stage designing. But paper sketches have their own disadvantage too. It lack of "designed memory"; hard to storage, arrange, search and reuse, especially lack of valid capability of interaction and alternation. For this reason, to study a sort of design tool that can combine paper-and-pencil hand-drawn sketch with computer is the hope of designers and is of great significance. Electric circuit diagram design is an impor-tant field of CAD technology.The thesis pointed against the defects of CAD technology to study thoroughly on-line recognition of hand-drawn electronic component symbol and make research and experiments on the key component of on-line hand-drawn electronic component symbol. The main part works as follows:1. A two levels hand-drawn electronic component symbol feature selection and extraction method that aimed at RBF(Radial Basis Function) neural networks is stud-ied in this thesis. Structural feature and relationship feature of hand-drawn electronic component symbol are defined in this thesis. Moreover, the definitions are applied in the feature extraction of hand-drawn electronic component symbol. Feature extraction and selection is the most crucial element in recognition of hand-drawn electronic component symbol, which affects directly recognition effect. This thesis used RBF neural networks which have the best approximation capability to recognize on-line hand-drawn electronic component symbol. So selecting feature which has large in-tra-class distance and small inter-class variance can achieve high recognition per-formance. Through observing and analyzing carefully the structure of hand-drawn electronic component symbol, the thesis found the invariable and separable traits of hand-drawn electronic component symbol, who aimed at RBF neural networks classi-fier. The traits are hand-drawn electronic component symbol is made up of some basic strokes sequentially and the basic strokes exist some constraint relation. The thesis used the term "structural feature"to define the basic strokes and used the term "rela- tionship feature" to define the constraint relation. In addition, the thesis used archi-tectural feature as the first level feature of the system of classification and recognition, then used relationship feature as the second level feature of the system of classifica-tion and recognition. Moreover, the thesis set aside binary decimal digit that started with 0 or 1 for structure feature and set aside binary decimal digit that started with 2 or 3 for relationship feature.2. The method of learned RBF neural networks'center of radial function is im-proved in this thesis. Common method of learned RBF neural networks'basic func-tion center is k-means algorithm. K-means algorithm is sensitive to initial clustering center, that caused the results of the algorithm are not accurate enough even can't converge. The improved method of learning center can overcome the shortcoming mentioned above, take full advantage of training sample, decreased effectively the effects of isolated point and clustering accuracy, and improved the clustering effi-ciency.3. A specific classification system that has two levels in series classifiers which is made up of RBF neural networks is proposed in this thesis. In pattern recognition, the design of classifier is a key technology too. By analyzing multistage classifiers inte-grated technology, the thesis proposed a specific classification system that has two levels in series classifier which is made up of RBF neural networks. The two levels classifiers both used RBF neural networks. The first level classifier is used to pre-classify and the second classifier is used to fine-classify. Pre-classification used one RBF neural networks and used the first level feature as his input feature vector. Fine-classification used three RBF neural networks and used the second level feature as his input feature vector. Then a recognition system is designed to verify the validity of those methods.