A Transfer Learning Method for Handwritten Chinese Character Recognition

The deep convolutional neural network has achieved high accuracy in handwritten Chinese character recognition (HCCR). Large scale handwritten data collection as well as labor-intensive labeling work are required to train an effective model suitable for various writing styles. Typically synthetic data are generated with data augmentation to alleviate the scarcity of labeled training data in real applications. However, the domain shift between synthetic data and real data results in unsatisfying recognition accuracy, bringing a significant challenge. A transfer learning method is proposed for synthetic-to-real handwriting recognition to alleviate the issue. In the framework, a proposed convolutional neural network is pre-trained with the synthetic data as the source model. Then, the source model is optimized with real samples of a specific handwriting style in an unsupervised manner. The effectiveness of the proposed transfer learning method is validated through systematic experiments on the public dataset as the target domain data.