Study On The Character Evolution Rules Of Oracle Bone Inscriptions Based On VGG-Siam Network

The oracle bone inscriptions(OBIs)dating back more than 3,000 years are the earliest known mature writing system in China.The decipherment of OBIs is essential to the research of the formation and evolution of Chinese characters and has become the window of probing into China’s ancient civilizations.In the studies of oracle bone,the recognition of OBIs has always been a heavy task.Through the persistent efforts of generations of researchers,about 1/3 of OBIs have been successfully recognized.To date,about 2/3 of OBIs are still unknown in their semantics.Due to the dating of characters,the recognition work is progressing hard.At the same time,by traditional recognition methods,such as the font comparison method,researchers need to compare the unrecognized OBIs with the known ancient characters such as bronze inscriptions,seal script and official script to recognize the OBIs.Although the traditional method of recognizing OBIs has high accuracy,it needs professional scholars and deep attainments,resulting in slow progress.Therefore,using Artificial Intelligence to assist OBIs is a method to improve the accuracy and efficiency of OBIs interpretation.With the development of OBIs research,the data on OBIs itself is gradually accumulating.Making full use of the massive information of OBIs to combine machine learning technology with traditional Chinese character interpretation methods is effective to break through the bottleneck of low efficiency of OBIs recognition research.Therefore,based on the Siamese network and the evolution from OBIs to regular script,the studies on OBIs,mainly include two parts in this work:The first part of the thesis mainly includes two aspects.Firstly,construct the dataset of Chinese character evolution.We have established a Chinese character dataset including OBIs,bronze inscriptions,seal script and official script.With each Chinese character’s evolution as a group,there are 972 groups and4860 samples.In the aspect of image preprocessing,gray processing is firstly carried out,using the weighted average method.Then the image geometric transformation,which adopts four methods: horizontal flip,rotation,translation,and distortion,and unifies the image size into 105×105 pixels to meet the requirements of OBIs interpretation with artificial intelligence technology.Secondly,an adaptive Chinese character evolution small sample preprocessing method is proposed,that is the adaptive filter Adafilter method.We introduce the local image base decomposition algorithm and black-box optimization algorithm into the algorithm.By calculating the local image bases of the dataset and linearly combining these local image bases with the optimization coefficients,the optimization of the preprocessing filter is transformed into a lowdimensional optimization.The experimental results show that the Adafilter filter maximizes the expected generalization performance as a preprocessing filter.In the second part of the thesis,the evolution rules of the five development processes of Chinese characters are studied.Firstly,we propose a Chinese character evolution prediction model for the VGG-Siam network.The network model is optimized in three aspects based on the VGG16 network and the Siamese network.(1)We define the number of character evolution and the length of font sequence based on the convolution layer to improve the recognition ability of learning features.For the character evolution dataset,we rely on variable structures to learn the feature vectors,and to realize the evolution rules of Chinese characters from OBIs to regular script.(2)Using the discrete attribute of the character evolution dataset,we define the character evolution distance formula according to higher stability and fast calculation speed.(3)Through the above sampling methods,we transform the problem into a binary classification problem and define the Chinese character evolution loss function,which can avoid the gradient dispersion in the gradient descent calculation.The results show a high F1 value at the VGG-Siam network,which is better than the ordinary Siamese network.The influence of the overall evolution law of Chinese characters is systematically studied.Firstly,the characters between adjacent character evolution periods are analyzed.In the prediction process,from OBIs to Bronze Inscriptions is 79.25%,Inscriptions to Seal Script is 75.24%,and from Seal Script to the Official Script is only 41.53%.While the accuracy rate from the Official Script to the Regular Script is80.13%.It can be seen from the experimental results that the accuracy rate from the Official Script to the Regular Script is the highest,and the accuracy from Seal Script to the Official Script is the lowest.Secondly,we analyze the four rules of the whole evolution period of Chinese characters: Simplification,Multiplication,Homograph and Erroneous transformation.In the whole prediction process of character evolution,the similarity of Erroneous transformation in the whole character evolution is the lowest.The research results can provide important ideas and data support for computer technology to decipher the semantics of OBIs.