Improving Off-Line Handwritten Chinese Character Recognition with Semantic Information

Abstract

Off-line handwritten Chinese character recognition (HCCR) is a well-developed area in computer vision. However, existing methods only discuss the image-level information. Chinese character is a kind of ideograph, which means it is not only a symbol indicating the pronunciation but also has semantic information in its structure. Many Chinese characters are similar in writing but different in semantics. In this paper, we add semantic information into a two-level recognition system. First we use a residual network to extract image features and make a premier prediction, then transform the image features into a semantic space to conduct a second prediction if the confidence of the previous prediction is lower than a threshold. To the best of our knowledge, we are the first to introduce semantic information into Chinese handwritten character recognition task. The results on ICDAR-2013 off-line HCCR competition dataset show that it is meaningful to add semantic information to HCCR.