Ancient Chinese Recognition Method Based on Attention Mechanism

Abstract

Characters and symbols play an important role of historical development and cultural transmission. Automatic ancient character recognition has become a meaningful and typical task. However, the existing recognition methods mostly focus on the detection and classification of modern Chinese, there are lack of the research on ancient Chinese, especially pre-Qin characters. And the methods are mainly computer graphics, topology, support vector machines (SVM) and convolutional neural networks (CNN), these methods lack attention to character features. Thus, based on ancient Chinese characters dataset of Tsinghua Bamboo Slips, the method proposed in this paper add attention mechanism to recognition algorithms to replace traditional convolution in order to improve recognition accuracy. Besides, we propose a data augmentation method specifically for character images, as much as possible without changing the writing form of Chinese characters. Experimental results demonstrated that our method has achieved a top5 accuracy of 99.98% which is higher compared with other methods.