Radical Region based CNN for Offline Handwritten Chinese Character Recognition

In recent years, deep learning based methods have been widely used in handwritten Chinese character recognition (HCCR) and greatly improved the recognition accuracy. However, most of the current methods simply employ famous networks like GoogleNet without fully embedding the specific features of Chinese characters. Taking structural characteristics into consideration, we propose a radical region network structure to represent the radical region information (For example left, right, top and bottom radical regions). In our study, the character feature is represented as global feature while the radical region feature is represented as local feature. The multi-supervised training method is also used to learn two kinds of feature at the same time. Experiment results show the proposed methods improve recognition accuracy of current models. The performance of the best model has been raised to 97.42% on ICDAR 2013 offline HCCR competition database which achieves the state-of-the-art result as we know.