Research Of Implementation Of Support Vector Machine In Handwritten Numeral Recognition

In the era of information, handwritten numeral recognition technology has significance of practice, which has a broad perspective of application in file retrieval, office automation, mail system, bank bill process, table record, and so on. This technology has its own complexity, which makes it difficult to implement the current recognition system. However, the widely demands for the implementation of this technology in real life make it very hot in this domain for all time.Currently, the main methods for handwritten numeral recognition include Decision Tree, Neural networks and support Vector Machines (SVM), since 1990s SVM was introduced as a very hot topic in this domain,. The training of the SVM is equal to solving the problem of quadratic programming of linear restriction, which makes the distance between the two super planes that separate the two kinds of pattern points in feature space maximum, and also can ensure that the result is the global optimal solution. Because The SVM classifier can absorb the deformation of handwritten characters, it has a better generalization performance.However, its basic usage is special in two-class problem, it is necessary to expand it for multi-class problem. This paper introduces some tradition methods such as OVA (One -versus-All), OVO (One-versus-one), DAG (Directed Acyclic Graph SVM) and propose a new SVM algorithms. The main idea of this algorithm is classifying multi-class as two-class, and does it until each single class can be separate by filters. The performances for training time, speed, training set of this algorithm are superior to traditional methods not only in theory analysis but practice implementation.