Combination of ResNet and Center Loss Based Metric Learning for Handwritten Chinese Character Recognition
Ruyu Zhang, Qingqing Wang, Yue Lu · 2017
Nowdays, deep networks have played a dominant role in the filed of handwritten Chinese character recognition(HCCR), and their recognition performance have reported to surpass human beings significantly. Recent research shows that the performance of deep networks has been further improved with the assistance of metric learning, since more inter-class and intra-class information can be captured. Center loss is a powerful metric learning strategy, and has no necessity of selecting sample units compared with other metric learning strategies. In this paper, we explore the effectiveness of the center loss based metric learning in boosting deep networks for the task of HCCR. By combining with the residual network (ResNet), an accuracy of 97.03% has been achieved on the ICDAR-2013 dataset, which is much higher than the deep networks without metric learning.