Research On OBI Recognition Based On Deep Convolutional Neural Network

OBI(Oracle Bone Inscriptions)is the earliest mature writing system discovered in China so far.It is the source of Chinese characters and the root of Chinese excellent traditional culture.At present,the study of OBI has entered the information age.At the opening ceremony of the international academic seminar to commemorate the 120 th anniversary of OBI discovery in Anyang City in 2019,OBI big data platform,"Yin Qi Yuan Wen" was officially released.This platform is OBI's knowledge sharing platform and is open to scholars around the world free of charge.At present,OBI resources are mostly pictures,which is not conducive to the data input,storage,retrieval,transmission,etc.of the platform.Therefore,as a member of the research team of the data platform,it is more and more essential to carry out the research on OBI character recognition.This paper mainly studies the recognition of OBI on rubbings.The traditional method of character recognition is mainly based on the framework of "data preprocessing + artificial feature extraction + classification recognition".The recognition rate of printed characters is high,but the recognition rate of handwritten characters is not high.OBI is handwritten characters from the Shang Dynasty,with many irregular shapes and few data samples.OBI on rubbings also has background noise,so it is difficult to identify them.In recent years,off-line handwritten character recognition technology based on depth learning has been fully developed.Deep neural network has strong image multi-level feature extraction ability and can describe data features of different levels of text images.In this paper,the deep convolution neural network is used to study OBI recognition,The deep neural network OBI recognition framework with better recognition ability is designed.An improved network model is trained by constructing OBI dataset.The experiment proves that the model can better express OBI character features.The main research contents of this paper include:(1)Constructing the OBI dataset for network training and testing.It takes one year to cut out the OBI images on rubbings determined by experts from the ten descriptions of OBI,cut out the OBI on the rubbings,and label them to form OBI rubbings dataset OBIS163.There are 163 classes of OBI in the data set,each class has 300 original pictures,250 of which are selected as training sets,and the remaining 50 original pictures constitute test sets.(2)Preprocessing the data set.The pretreatment of data includes data enhancement,image denoising,and normalization.Among them,data enhancement adopts the method of image geometric transformation,including rotation,deformation,scaling,masking,etc.Image denoising uses the feedforward neural network denoising method and constructs a denoising neural network to denoise OBI data.Finally,the zero-mean normalization method is used to normalize it.(3)Construction of OBI recognition network based on deep learning technology.Firstly,four OBI are selected to carry out recognition experiments on OBI-CNN.The results show that the Top-5 recognition rate of the test set is only 70.71%,so a network model is proposed,named OBI-CNN,which can give consideration to the recognition speed and accuracy of OBI.The model is designed according to the characteristics of OBI.Because OBI are carved with sharp tools,the fonts are mostly strips,not squares of Chinese characters.Therefore,part of the square convolution kernels are replaced by strip convolution kernels,and the characteristic graphs of the two strip convolution kernels are superimposed to deepen the depth of the network and reduce the number of network parameters at the same time,making the characteristics of OBI more obvious.The experimental results show that the improved network can better extract the features of OBI characters,and the recognition rate reaches 84.45%,which is 13.74% higher than the network before the improvement.(4)The establishment of the OBI character recognition system.Taking PyQt5 as an interface development tool,based on the deep learning framework Pytorch,the network model trained is used as an OBI depth feature extractor by integrating the relevant algorithms of different links such as OBI feature extraction,classification,and recognition,and is loaded on the Windows system to make OBI recognition system convenient for users to use.