The Design Of SVM-based On-line Handwritten Recognition Classifier

Although on-line handwritten recognition has been developing for decades, its effect still doesn't come up to expectation. However, in recent years, along with the emergence of keyboard-less electronic products (smart phone, tablet, etc.) and the rapid advance of MEMS and image processing, handwritten input has attracted more and more customers and in the meantime new research and application fields have appeared, such as space handwritten recognition, gesture identification, signature ver-ification, mathematical equation recognition, chemical symbol recognition, and so on. All these facts lead to an increasing interest in the study of on-line handwritten recog-nition.Support vector machine is a new pattern recognition algorithm and is developing fast. It is based on statistical learning theory, kernel method and generalization theory and adopts the principle of structural risk minimization to compute the optimal sepa-rating hyper-plane. Compared with other pattern recognition algorithms, SVM has a substantial theoretical foundation. In the field of pattern recognition (speech recogni-tion, genetic testing, handwritten recognition, etc.) and its related fields (state predic-tion, curve fitting, etc.), SVM enjoys reasonable recognition effect.This thesis studies Gaussian Dynamic Time Warping kernel, which is applied most successfully in SVM-based on-line handwritten recognition classifier design, and its two shortages:it is designed for a variety of pattern recognitions and thus compared with other algorithms the advantage of SVM hasn't been fully utilized; its computation complexity is high and thus the running time is comparatively long. To solve these two problems, this thesis studies the characteristic of feature vectors, pre-sents a method of optimizing GDTW kernel and explores the affection of different calculation ways of the optimal alignment path.In order to verify the validity, the thesis designs a SVM classifier on the basis of optimized GDTW kernel and conducts an experiments with on-line handwritten recognition database UJIpenchars2 (with high quality) and UNIPEN (with low quali-ty). The result shows that the method reduces the number of support vectors and in- creases recognition efficiency.