Research On Theory And Applied Technology Of Handwritten Chinese Character Recognition In HCL2000

Since handwritten Chinese character recognition appears a tough job, researchers are dedicated to higher recognition rate through improvement on methods. In fact, it is not fair to make a judgment whether method is effective or not only by the criterion of recognition rate. This paper is on the subject of theory and applied technology of off-line handwritten Chinese character recognition, which focuses on the development of evaluation system of theoretical methods. An optimization model of similarity/dissimilarity measure is proposed to discuss on characteristics of discriminating of dimensionality reduction techniques such as LDA, LPP and MFA. Non-linear methods describe relation between sample pairs, while linear techniques represent spatial distribution of data. In general, the former employing gradient feature far outperforms the latter. In this paper, human performance on handwritten Chinese character recognition is compared to machine, which aims to present a criterion baseline of required accuracy for further research. HCL2000, one of the largest databases of handwritten Chinese characters, introduces sample into evaluation. According to experiment results, the best machine record is competitive with average human performance. A system of Human Performance Test on CAPTCHA is designed to verify the security of identification code in the form of handwritten Chinese character images. With extremely high error rate, OCR software barely recognizes distorted character images of noise, lines, colors, etc., which proves this instance of CAPTCHA is safe in practical applications.