Printed Japanese character recognition based on multiple modified LVQ neural network
Koji Miyahara, Fumio Yoda · 2002
A multiple modified LVQ neural network model that can recognize Japanese characters over 3000 categories with high performance both in accuracy and speed is proposed. The multiple modified LVQ network is based on the LVQ (learning vector quantization) neural network and a large scale of network can be implemented easily because of its simple structure. This network has a training function of fast convergence and of easy modification without disturbing past trained weights for Japanese character recognition. An experimental system using a neuro-computer with four digital neuro-chips and experimental results are described. With the experimental system it takes 18 minutes to learn 35,000 samples by 20 training cycles, while it takes more than one week with a workstation. Moreover it can recognize about 350 characters a second for 3584 categories. High recognition rate of 100% for training fonts and of over 99% for testing fonts were achieved with 49,500 samples.>