Deep Learning and Image Processing Combined Organization of Shirakawa’s Hand-Notated Documents on OBI Research

Abstract

The purpose of this work is to organize Professor Shirakawa’s newly discovered hand-notated documents on his Oracle Bone Inscription (OBI) research. During the second half of the 20th-century, Professor Shirakawa was a prominent researcher on Chinese culture, especially in the field of OBIs, and he left behind many research documents he had notated by hand. However, some of these documents have not been properly organized yet. The reorganization of OBIs is not only helpful for better understanding his research but also for further studying about OBIs in general and their importance in ancient Chinese history. Part of Professor Shirakawa’s hand-notated research documents on OBIs is introduced to the world for the first time in this work. For organizing these documents, Firstly, a morphology-based segmentation method is applied to segment the characters in the documents and then the paper proposes a slight neural network for removing the noise from the mis- segmented characters. Finally, a dynamic K-means method is applied for classifying the segmented characters. Specifically, the histogram of oriented gradients (HOG) descriptors are extracted as features, and the class number of K is dynamically decided by using the silhouette coefficient. The results of this evaluation showed that the accuracy of noise and character classification after segmentation achieves 96.50%, and the accuracy of character classification achieves 74.91%. The results demonstrate the effectiveness of the proposed method.