Research On Bronze Inscriptions Character Recognition Based On Deep Learning

Bronze inscriptions are the main carriers to record and communicate during the Shang and Zhou Dynasties in ancient China.It plays an important role in the study of Shang and Zhou Dynasties’ history and community life.A basic work on bronze inscriptions is its character recognition.However,this work is still completed by experienced paleographers manually.Computer image recognition technologies,particularly deep learning methods,have achieved great success in the character recognition task.These methods could help computer to recognize characters in images quickly and efficiently.Based on the above demand and background,this paper aims at using deep learning methods to complete the bronze inscriptions character recognition task.The main work of this paper is as follows:(1)The collection and organization of a new Shang and Zhou bronze inscriptions characters dataset.We collect and obtain a new formatted characters dataset and implement data augmentation based on the dataset.The general methods of data augmentation,such as translation and rotation,is used to expand the dataset.The secondary development of Rhino 3D software is proposed to simulate lighting effects on character images in the dataset.Implicit semantic data augmentation(ISDA)is introduced to further achieve data augmentation,which can help regularize deep neural networks model to obtain stronger generalization ability.(2)The research Shang and Zhou bronze inscriptions character recognition based on attention mechanism.This paper researches deep neural networks for bronze inscriptions character recognition.We explore many advanced network architectures such as VGG and Res Net and obtain good result.Then,the spatial attention mechanism is introduced into these networks by using a spatial transformer network(STN)after the data input layer.It makes the whole system achieve better recognition accuracy.(3)The improvement to the sample unbalanced problem by using image generation methods.The Shang and Zhou bronze inscriptions characters dataset has the unbalanced problem in inter-class and inner-class samples which affects the performance of character recognition deep neural networks.This paper uses generative adversarial network(GAN)and variational auto-encoder(VAE)methods to generate needful and useful bronze inscriptions character images which could be expanded into the character training dataset.The final result of experiments confirms the effectiveness of these generated character images.A number of experiments with detailed analysis have been done and finally we reach 91.387% accuracy showing the feasibility and effectiveness of this work.