Abstract
The music notation of Guqin (“古琴”, Chinese seven-string zither) named “Jianzi Pu (減字譜, simplified music notation of Guqin)” was invented at the Middle of A.D. 700, and Guqin music remained more than 600, however, only about 100 of them are played in nowadays. The reason is that the handwritten “Jianzi Pu” is hard to be understood even for experts or professional Guqin players. In this study, we applied deep learning methods such as VGG16 and YOLOv5 to the recognition of a Guqin notation “Sen-O-So” (仙翁操, Melody of the Immortal Elder). Firstly, we created a dataset including 55 kinds of single characters of Sen-O-So in 4,951 images from 23 versions found on the Internet and obtained by data augmentation, i.e., image processing such as rotation, enlargement (zoom-in), reduce (zoom-out), various filtering, etc. Secondly, we compared the recognition rates of VGG16 and YOLOv5 in the experiment. The average accuracies of 55 classes images by VGG16 and YOLOv5 were 87.50% and 88.47% respectively for the test data. Additionally, we created a dataset of Sen-O-So video clips to match the recognition results of single characters by YOLOv5 and realized an online ancient music restoration system development.
https://ror.org/05h68bp56 
