A scheme of on-line Chinese character recognition using neural networks
Hwei-Jen Lin, Shwu-Huey Yen · 2002
The paper proposes a scheme of online Chinese character recognition, based on neural networks. The supervised backpropagation algorithm is used to train the network. The input character is converted as a sequence of virtual stroke segments as well as real stroke segments, which is a good feature exactly describing the complete structure of a character, and is to be extracted by our system. In order to simplify the recognition process and reduce the recognition time, the neural network is divided into several subnetworks. Each of them is responsible for recognizing a group of about 75 character patterns. In other words, the huge set of Chinese characters is divided into several groups according to the numbers of stroke segments in the characters, and for each group of characters, a specific subnetwork is trained in order to recognize every character in the group. Whenever the system accepts an input Chinese character, it will calculate the number of stroke segments, including virtual stroke segments as well as real stroke segments in that character, and then determine which subnets to enter for recognition process. The system is allowed to accept and recognize some interconnected characters. The algorithm was experimentally implemented in a personal computer system, it accepts interconnected Chinese characters written on an electronic tablet, and performs recognition in real time. Our experiment showed that recognition accuracy exceeded 96% on the test example.