| Feature extraction is a key part of pattern recognition system. After analysis the existing methods of feature extraction, this paper researched the process of feature extraction for on-line handwritten Tibetan character recognition. During the feature extraction process, the paper firstly extracted the statistical and structure features of Tibetan character, including geometric features,directional features and the projector distance measure features. After that, the directional feature of Tibetan character handwriting was analyzed, compared and counted in four, six and eight directions which influenced the recognition rate. Meanwhile, the structure of Tibetan characters both on global and edge was divided, statistical features of each cell in the domain of Tibetan characters was counted, the character recognition rate via the different division was compared. Secondly, this paper fused the features that extracted by different methods to form the new feature set. Finally, the new feature set of feature was selected by FEND (Feature Extraction No Dimensionality),LDA,2DPCA + LDA methods respectively.Experiments used 315 sets of training samples and 70 sets of test samples, and each sample consisted of 562 Tibetan characters. The results showed that: six directions are better than four and eight directions in the area of feature extraction. 80 blocks increases the recognition rate in the division of character structure compared with 40 blocks and 60 blocks. In the case of increasing feature set, the top ten average recognition rate of feature extraction based on statistical and structure of on-line handwritten Tibetan character is 90.95%, and 2DPCA + LDA over LDA method improves on-line handwritten Tibetan character recognition rate, compared with the LDA method, the recognition rate increased by 1.96%. |
