A model based detecting approach for feature extraction of off-line handwritten Chinese character recognition

Abstract

To simulate the cognitive ability of a human brain, especially thinking in images, a dynamic network composed of model based detectors for feature extraction of off-line handwritten Chinese character recognition (HCCR) is proposed. It is first argued that, according to noetic science, the methods of HCCR can be divided into two categories: thinking in images and thinking in logic. The former is particularly emphasized. A multilayer representation of an attributed semantic network for Chinese characters is given. After that, how a model based substructure detector makes stroke segmentation, detects points, strokes and their relationships, and extracts features is put forward. Finally, the dynamic association procedure of the model based detector network which is composed of 1700 substructure detectors is also explained. A Chinese character recognition system has been built based on this feature extraction approach. >