Robust coefficient selection for recognition of complex Chinese characters

In this paper, two fast and robust methods, maximum non-overlapping interval (MNOI) and vector space (VS) methods, are developed for the selection of the best necessary transformed features for the recognition of machine printed Chinese characters. The first method generates a binary decision tree while the second generates an M-dimensional vector representing the character and recognition is achieved by finding the minimum Euclidean distance. Characters are preclassified and then transformation is applied before the final recognition stage. Noisy characters are used to test the performance of the classifiers in terms of speed and accuracy. Using a 33 MHz 386 PC, our experimental results show that the VS method achieves an overall recognition rate of 99.93% for characters with SNR better than 20 dB while the MNOI method achieves a recognition speed of 200 chars/sec. Furthermore, it is found that the Discrete Cosine Transform is the best transform core compared with Haar and Walsh Transforms