Research Of Oracle Bone Inscription Detection Based On Deep Convolutional Neural Network

Oracle Bone Inscription(OBI)is the earliest systematic writing found in China and the only language among the four ancient civilizations that have been passed down and influenced so far.On October 30,2017,OBI was inscribed on the UNESCO Memory of the World Register,which marks that their value has been widely accepted in the world.On November 1,2019,President Xi pointed out that OBI is the source of the Chinese writings and the root of Chinese civilization that should be treasured,inherited,and cultivated in the congratulatory letter to Oracle Bone discovery and research 120 th anniversary.OBI detection is the first study in this field.The purpose of detection is to locate characters from OBI rubbings,which is the basis of the following tasks such as recognition,retrieval,and decipherment.Currently,these works need to be implemented manually by experts,which is inefficient.Among the existing 160000 pieces of oracle bones approximately,only 80000 have the explanatory document,but the subjective cognition of experts also have been attached,which is inaccurate.How to precisely detect and recognize OBI characters is still an urgent task.Given the success of deep learning in computer vision,this paper carries out relevant research of deep convolutional neural networks on OBI detection,and the main work is as follows:(1)A large scale of the OBI detection dataset is collected,which contains 9400 samples,and is still being updated to serve relevant research.(2)Secondly,mainstream object detection models including Fast R-CNN,FPN,SSD,YOLOv3 and Retina Net are transplanted to the OBI detection dataset.The experiment shows that YOLOv3 holds the best overall performance.(3)SPPG-YOLO and ASPP-YOLO are proposed to improve the performance of YOLOv3.The former uses a series of pooling layers to integrate features and improve the detection accuracy,but only brings a little bit of computational overhead;the latter takes the module of semantic segmentation branch into the detection task.Although it is not as fast as the former,the detection result is more accurate.(4)According to the feature of OBI,multiple optimizations are carried out,including noise simulation,anchor box clustering,and GIo U loss function.The experiment shows that the proposed model combined with these optimizations achieves the state-of-the-art in the OBI detection task,and the F1 score reaches83.4 %.(5)Finally,models studied in this paper are packed into software,and affiliate Efficient-Net in the tail,implementing an end-to-end OBI detection and recognition system.