Oracle Graphics Recognition Based On Capsule Network And Transfer Learning

As the earliest mature text discovered in China,Oracle records the economic and political situation of the Shang Dynasties,reveals the social appearance of ancient China,and is significant in cultural and academic research.With the rapid development of Deep Learning algorithms,the application of Image Recognition methods to solve problems of the Oracle is becoming the direction of the future.Studies have shown that Oracle has a relatively complete Construction system,and each Oracle Graphic is composed of one or several Oracle Radicals according to specific rules.The Construction system is the key to the study of Oracle and has high research value.However,it is challenging to apply the Construction rules to the task of computer recognition of Oracle.In order to solve the problem,in this thesis,the construction system of Oracle has been carefully studied.Then,a new recognition method of the Graphics,which is based on Deep Learning and uses the knowledge of the construction rules,is proposed.In the end,its effectiveness was verified by experiments.Firstly,in this thesis,an in-depth study on the Graphics and Radicals is conducted,and the characteristics of the Construction system has been discussed.,Two large data sets were created for Radicals' feature extraction and Graphics recognition tasks,and they are Radicals' data set Radical-148 and Graphics' data set Oracle-250.Secondly,RadicalNet,the model based on Capsule Network and Transfer Learning,is designed to recognize and extract features of Radicals.The model is trained by the Radical-148 dataset to perform multi-object recognition of Radicals contained in the Graphics.In the experiment,the optimal model achieves 68.73% Top2 accuracy and 75.02% Top5 accuracy on the Oracle-250 dataset.The result verifies the validity of RadicalNet in the task of feature extraction and recognition of Radicals.Finally,in this thesis,two Oracle Graphics recognition model,OracleNetV1 and OracleNetV2,based on Transfer Learning was proposed.The models are based on the Radicals features extracted by RadicalNet and finally,recognize the classes of the Graphics.In the final experiment of Graphics recognition,on the test set of Oracle-250,OracleNetV1's Top1 and Top5 accuracy reached 87.04% and 91.18%,respectively,and OracleNetV2 reached 90.73% and 93.79%,respectively.It is verified that the OracleNet models have powerful and robust oracle recognition abilities.