Off-line handwritten chinese character recognition by using support vector machines

Abstract

The research proposes an efficient high-accuracy SVM-based handwritten Chinese character recognition system. The 304-dimensional feature vector of a handwritten Chinese character consists of the contour direction feature, the crossing count feature, the peripheral background area feature, and the contour line length feature. The mean-vector recognition method is first used as the coarse classifier to get a small number of candidate classes for the input vector. After the preliminary multi-class SVM-based recognition method is trained by 200 instances per candidate class, the recognition rate for the test handwritten characters of 5,401 classes can achieve 98.31%, which is much higher than about 93% accuracy for the mean-vector recognition method alone. To speed up the recognition, the forward and the backward greedy two-class SVM-based recognition methods are proposed to greatly reduce the recognition time to a practical level.

Subject Classification:

• 94A08
• 68T05
• 68T10

Keywords:

• Support vector machines (SVM)
• Coarse classification
• Handwritten Chinese character recognition
• Candidate class