Offline recognition of handwritten Chinese characters using Gabor features, CDHMM modeling and MCE training

Abstract

We've been developing a Chinese OCR engine for handwritten Chinese scripts. Curre ntly, our OCR engine supports a vocab ulary of 4616 characters which include 4516 simplifi ed Chinese characters in GB23 12-80, 62 alphanumeric characters, 38 punctu ation marks and symbols. By using 1,384,800 character samples to train our recognizer, an averaged character recognition accuracy of 96.34% is achieved on a testing set of 1,025,535 character samples.An arguably best Chinese OCR product on the market achieves an accuracy of 94.07% for the recognizable Chinese characters in the above testing set. In this paper, we describe key techniques used in our recognizer that contnb ute to the high recognition accuracy, namely the use of Gabor features and their spatial derivatives as raw features. the use ofLDA for feature extraction and dimension reduction, the use of CDHMMs for modeling Chinese characters along both horizontal and vertical directions, and the use of mini mum classification error as a criterion for model training.

Offline handwritten Chinese character recognition by radical decomposition

“…We know of only one other study of Chinese character recognition that has produced comparable results-that of Ge et al [2002], whose approach is considerably different from ours. They use four Gabor features (see Lades et al [1993]) and their spatial derivatives to parameterize the whole character image in both horizontal and vertical directions.…”

“…In this paper we have presented just such a "new way," which we believe has many potential advantages over the work of Ge et al [2002]. Their approach is effectively holistic, encoding the complete character image, which implicitly carries the disadvantage of trying to model simultaneously the three main sources of variability in Chinese handwriting, which we outlined in Section 3.…”

“…Gabor filters are spaced every 8 pixels, and the CDHMMs are trained using a minimum classification error criterion on 1,384,800 character examples. For a set of 4,516 character classes, Ge et al [2002] obtain a recognition rate of 96.34% on a test set of 1,025,535 characters. Superficially, they obtain a better recognition rate on a larger test set with a larger number of character classes.…”

“…It is not clear that Ge et al's results are necessarily any better than ours; it is safest to say that they are comparable. Also, as Ge et al [2002] say, "offline recognition of . .…”

“…. handwritten Chinese characters remains largely an unsolved problem," a fact which encourages the search for "new ways to attack this difficult problem" (p. 1053Ge et al [2002). …”