Data Augmentation Based on CycleGAN for Improving Woodblock-Printing Mongolian Words Recognition

In order to improve the performance of woodblock printing Mongolian words recognition, a method based on cycle-consistent generative adversarial network (CycleGAN) has been proposed for data augmentation. A well-trained CycleGAN model can learn image-to-image translation without paired examples. To be specific, the style of machine printing word images can be transformed into the corresponding word images with the style of woodblock printing by utilizing a CycleGAN, and vice versa. In thisway, new instances of woodblock printing Mongolian word images are able to be generated by using the two generative models of CycleGAN. Thus, the aim of data augmentation could be attained. Given a dataset of woodblock printingMongolian word images, experimental results demonstrate that the performance of woodblock printing Mongolian words recognition can be improved through such the data augmentation.