Recognition of Oracle Bone Inscriptions Using Deep Learning based on Data Augmentation

Abstract

Oracle bone inscriptions are among the oldest kind of characters in the world and were first inscribed on cattle bone or turtle shells about 3,000 years ago. They were discovered in 1899, and unfortunately very few papers described them. Moreover, the aging process has made the inscriptions less legible. Understanding the inscriptions is important in terms of researching world history, character evaluations, and more. This work introduces a state-of-art initiative to recognize oracle bone inscriptions by deep learning. This is the first time an oracle bone inscription dataset featuring real rubbing images has been generated. For training before recognition, we augment the inscription images by means of rotation, Gaussian noise addition, cutting, brightness changing and inversion, which turns one image into 3,072 new images. We change the dropout, layer number, and filter number of every layer, so as to improve recognition ability and achieve a prefect recognition rate. By analyzing the mistaken recognitions, we identify some special characters that should be given special data augmentation. In an experiment using 184 difference characters and a dataset consisting of 2,000 images including 538 test images, parameter tuning and data augmentation resulted in the recognition rate of 92.3%, and the training data was augmented into 9.69 million items.