Multi-stage Feature Extraction in Offline Handwritten Chinese Character Recognition

Abstract

Convolutional neural network (CNN) has achieved tremendous success in handwritten Chinese character recognition (HCCR). However, most CNN-based HCCR research nowadays focus on complicated and deep CNN module, rarely analyzing the whole feature extraction process which has a crucial impact on the final recognition rate. In this paper, the following two questions are answered: (1). Information loss is inevitable on the training stage of complex learning problems, but at which layer does the information loss mainly occur; (2). Different layers have different effects on CNN, what is the best place for multi-stage feature extraction that influences CNN most. We make use of the proposed module in typical CNN and analyze classification results on CASIA-HWDB1.1. It is shown in this paper that, (1). Multi-stage feature extraction achieves better performance on HCCR than single stage feature extraction. (2). Multi-stage feature extraction should be designed at the convolution layer rather than the pooling layer. (3). Multi-stage feature extraction designed at shallow layers outperforms that designed at deeper layers. By analyzing the structure of multi-stage feature extraction, we propose an appropriate CNN approach to HCCR, which achieves a new state-of-the-art recognition accuracy of 91.89 %.