Thangka Image Classification Based On Deep Learning

Thangka is a kind of scroll paintings created on specially made curtains or paper.There are many themes,such as Buddha Thangka with Buddhism as the theme,biography Thangka to depict characters and events,and historical Thangka to record,etc.These images record rich history and culture and have high research value.In the study of thangka,the distinction between the main and the main categories is the basis for the knowledge mining of thangka domain.In recent years,with the emergence of large tagged data sets and efficient computer processors,deep learning has been rapidly developed,and many excellent network models have been proposed.Among these network models,Convolutional Neural Network has achieved high performance in many image classification methods because of its deeper network structure and a large number of network training parameters.In this thesis,convolutional neural network is used to classify three kinds of Thangka Buddha,Vajra and Sakyamuni.Specific research contents include:(1)Thangka has complex color and texture features,and the background of some images is very similar.Traditional classification methods are difficult to extract effective features,so the classification effect is not good.Based on the Dense Net network,this paper designs the Thangka classification network model.The concrete measures include: first,the compression-excitation module is added,and the addition operation is introduced to replace the multiplication operation;Secondly,a weight calibration method combining global maximum pooling and global average pooling is designed to recalculate the weight value of the image feature image.Finally,the experimental analysis on the Thangka image dataset shows that the addition operation not only speeds up the network training speed,but also improves the performance of the model,and the classification accuracy reaches 95.51%.(2)In the process of network training,the recognition rate of the category with relatively small sample size is much lower than that of the category with relatively large sample size,which greatly reduces the network performance and results in a great deviation of the classifier obtained.In order to overcome this problem,this paper proposes a Thangka classification network model.Firstly,the weight parameters of the first part of the network layer are frozen during network training to update the weight parameters of the second part of the network layer due to the great difference between the Thangka image and the natural image.Secondly,the loss value of the model is calculated by combining the cross entropy loss function and the complementary cross entropy loss function.Finally,the experimental data are obtained by comparing with the Res Net network,and the results show that this method has a better classification effect.The classification accuracy of Thangka main and respected classification reaches 94.57%,which is 2.3% higher than that of Resnet50 network.(3)A Thangka classification system is designed and implemented,which consists of six modules: login,registration,classifier,user management,label management and configuration file management.The proposed Thangka classification model is deployed to the system to realize the automatic classification function of Thangka classification module.