Oracle bone inscription is the earliest writing system in China which contains rich information about the history of Shang dynasty. Automatically recognizing oracle bone characters is of great significance since it could promote the research on history, philology and archaeology. The proposed solutions for oracle bone characters recognition are mainly based on machine learning or deep learning algorithms which rely on a large number of supervised training data. However, the existing dataset suffers from the problem of severe class imbalance. In this work, we propose a CycleGAN-based data augmentation method to overcome the limitation. Via learning the mapping between the glyph images data domain and the real samples data domain, CycleGAN could generate oracle character images of high-quality. The quality is evaluated using the quantitative measure. Totally, 185362 samples are generated which could serve as a complementary to the existing dataset. With these generated samples, the state of the art results of recognition task on OBC306 are improved greatly in terms of mean accuracy and total accuracy.