Oracle Bone Inscriptions Recognition Based on Deep Learning

The research and decipherment of oracle bone inscriptions have been in progress, however, due to the incompleteness of a large number of topographies, the wide range of meanings of the characters, and the complexity of the principles of character creation, the decipherment of oracle bone inscriptions has been slow, and it requires a lot of time to be spent on the decipherment by professionals. In this paper, we try to use the deep learning method to automatically recognize some deciphered single words, in order to reduce the workload of professionals and makes it easier to recognize and decipher oracle bones.This paper divides oracle recognition into two main phases: the detection phase and the recognition phase, the first stage inputs the picture into the detection model, selects the region that may be the oracle bone single character, and then sends it to the recognition model in the second stage, and the recognition model determines which single character is the character in the region. Three detection models, YOLOv5(You Only Look Once version 5), FAST-RCNN (Fast Region-based Convolutional Neural Network)and DETR(DEtection TRansformer), are trained in the first stage, and two models, YOLOv5-classify and ResNet(Residual Network), are trained in the second stage, in order to horizontally compare the performances of the different models on the same task, and select the optimal model to complete the oracle bone recognition task. Eventually, the best performing detection model on the dataset is YOLOv5, which achieves an average accuracy of 98.8% under the threshold of intersection and concatenation ratio of 0.5; and the best performing recognition model is ResNet, which achieves an average accuracy of 99.8% for the detection of individual single words with an F1-score of 0.849.