Research On Painting Image Style Classification And Application Based On Deep Learning

With the continuous improvement of living standards,it has become a trend for people to constantly seek to improve their cultural literacy,such as visiting art exhibitions and appreciating paintings by famous artists.However,how to find the style I am interested in from a large number of paintings and understand the style characteristics of the works is an urgent problem to be solved.This paper proposes two style classification models of painting images based on deep learning and applies them in the field of Thangka painting.The main contributions of this paper are as follows:(1)A painting image style classification model based on Res Net is proposed.Based on Res Net50,this paper introduces attention mechanism into the model,and then optimizes the activation function of the model.The deep network model naturally has not translational invariance,so the model is improved to solve this problem.and the recognition accuracy of the improved model on Pandora data set is 69.4%.(2)The NTS network model is improved.Blur pool is combined with the feature extractor module of NTS network Secondly,the learning rate reduction strategy of the original NTS network model is improved.The cosine annealing attenuation strategy is introduced into NTS network model to further improve the accuracy of painting image style classification.The improved NTS network model was able to identify the Oil Painting data set with an accuracy of 73.8%.(3)A thangka school data set is constructed.In this paper,the specific production process of the data set is introduced,which mainly includes Niboer painting school,Miantang painting school and Gamagazi painting school.Prior to this,there was no thangka data set based on the classification of painting schools.(4)The generalization of the improved model was verified by using a self-made Thangka school data set.The recognition accuracy of the improved models on the Thangka school data set is 87.4% and 95.1% respectively,which verifies the generalization of the two schemes proposed in this paper.