Research And Application Of Oracle Bone Inscriptions Recognition Based On Convolution Neural Network

Oracle Bone Inscriptions is the earliest mature writing system in China.It is the origin of Chinese characters and the foundation of the excellent traditional culture of the Chinese nation.The task of oracle bone inscriptions recognition is to determine the character category of oracle bone inscriptions,which is a necessary prerequisite to complete the interpretation of oracle bone inscriptions.At present,the automatic character recognition of oracle bone inscriptions on rubbings largely depends on experts’ feature engineering,which is a great laborcosting and time-consuming job with complicated work content and heavy workload.Therefore,automatic image recognition of oracle bone inscriptions has important research value.At present,by means of convolutional neural networks,substantial achievements have been yielded in the image recognition domain.This paper mainly studies oracle character recognition by using the ResNet-50 network model in the convolutional neural network.Through changes of convolution kernel dimension,adjustment and optimization of model parameters and algorithms,this paper designs an oracle bone character recognition model with better recognition capacities conducts contrast experiment on this model with five other classic convolutional neural networks.It is provided by the experimental results that the model with high recognition accuracy has strong capabilities of oracle image recognition.The research in this paper consists of following contents:First,oracle data set used in the experiment is preprocessed.Among 300 oracle text categories of original oracle data set,a total of 36,000 pictures are assigned uniformly,120 pictures for each category.Data preprocessing involves data set expansion,image denoising and image size normalization.Image cropping and image geometric transformation methods are adopted for data set expansion.The image geometric transformation is achieved through rotation,translation and elastic deformation.The Attentive Generative Adversarial Network is used for oracle image denoising by establishing a denoising network model..Image size normalization uses bilinear interpolation method,which adjusts all the image sizes in the data set to 224 * 224 pixels.After expansion,the number of images in each image category is 960,which is 8 times of the original data set.Secondly,this paper proposes an improved ResNet-50 network model.Considering the rectangular shape of oracle characters,in the model,the size of some convolution kernels in the convolution layer of the original model is changed from the square convolution kernel to rectangular convolution kernel.Then this paper studies the optimization of the ResNet network model,compare the optimization effects of each method on the model through experiments through controlled variable method,so as to finalize the optimal plan.It is determined that the final solution sets the batch＿size value as 64,adopts Adam optimization algorithm through piecewise constant learning rate attenuation method.The recognition accuracy of the optimized model is 87.43%,which is 7.38% higher than that of the original model.Finally,the improved network model is compared with other five classical convolutional neural network models.The experimental results show that the "improved ResNet-50" network model proposed in this paper has the highest recognition accuracy.Thirdly,an oracle recognition system is designed and implemented by summarizing previous work,including image denoising algorithms,image recognition algorithms.Oracle pictures uploaded by the user to the system can be denoised in virtue of the well-trained image denoising network in the system.Then,they will be input into the trained "Improved ResNet-50" image recognition network for recognition,and recognition results will be finally output onto the front-end page.