Research On Dynasty Identification Method Of Ancient Murals Based On Deep Learning

In mogao Grottoes,there are a large number of mural grottoes with distinctive characteristics of dynasties,which have developed and continued in the timeline of multiple dynasties.The mural paintings of each dynasty have their own characteristics of the times in terms of figures,colors and painting characteristics.The key step in the further study of these rich murals is to identify the dynasty of the images.This thesis takes chinese traditional mural images as the research object,aiming at the problems such as the ambiguity of mural images themselves,the difference of painting styles of mural paintings in different dynasties,and the time-consuming and laborious problems of the traditional manual direct reference of mural texts or historical documents to identify the dynasties of mural paintings.In order to solve the above problems,the main research methods and principles of this thesis are as follows:1.An ancient mural dynasty classification model incorporating transfer learning.First,in order to overcome the difficulty of collecting mural image datasets,the mural images are quantitatively expanded through various data enhancement techniques,and the data-enhanced mural images are used as the main dataset for the experiment.Then,using Inception-v3 as the basic model framework,transfer learning is integrated into the model.Finally,the color histogram and LBP texture histogram are introduced to extract the artistic features of the mural images,and the contextual information of the mural images is connected to construct an ancient mural dynasty classification model that integrates transfer learning.2.An algorithm for ancient mural dynasty identification by adaptive enhanced capsule networks.Based on the basic structure of the original capsule network,this thesis proposes an adaptive-enhanced capsule network for ancient mural dynasty recognition algorithm.First,increase the number of convolutional layers,and change the original single convolutional layer to three consecutive activation layers to perform high-level feature extraction on mural images.Then,perform uniform layer activation on all feature extraction layers;At last,in order to make capsule network is more suitable for feature extraction of mural images.The feature capsule layer parameters and category capsule layer parameters of capsule network are adjusted according to the characteristics of mural images.Finally,the research on the dynasty recognition algorithm of ancient murals is realized.To sum up,this thesis has completed the construction of the mural data set,the construction of the Inception-v3 model integrating transfer learning,and the realization of the ancient mural dynasty recognition algorithm of the adaptive enhanced capsule network.The experimental results show that the model and algorithm proposed in this thesis can effectively identify the dynasty to which the mural belongs in the identification of ancient mural dynasties,and has a good classification result.