Research On Super Resolution Reconstruction And Classification Optimization Of Murals Based On Deep Learning

Ancient Chinese mural paintings once had a glorious history.After thousands of years of accumulation and accumulation in dynasties,there have been art classics like Dunhuang murals that are still breathtaking today.Recreating the old style of the work has become the basis of research.Due to the differences in the painting styles of various types of murals and the harsh geographic environment where the murals are located,the classification of murals is difficult and the texture details are ambiguous,which has become an obstacle to exploring the value of murals,and there is still a risk of damage to the murals after manual restoration.Continually discovering and protecting the artistic value of murals is not only the direction that mural researchers are pursuing,but also the goal of this thesis.With the continuous development of computer technology software and hardware,we have found a way to intelligently protect the artistic value of murals,allowing computers to "learn" autonomously,with the ability to intelligently classify and preventively protect digital murals,that is,to classify and super-resolve digital murals.Rate reconstruction.Applying the deep learning method to the digital mural scene,intends to solve the problem of mural classification and preventive protection,and provides a new solution to the problems encountered in the protection of murals.(1)Stable enhancement generates super-resolution reconstruction model of countermeasure network.The model is improved based on the generated confrontation network,in which dense residual blocks are used to extract image features.At the same time,three loss joint optimization models,namely perceptual loss,content loss and confrontation loss,are introduced.When calculating the perceptual loss,the characteristic information before activation is used for calculation.The confrontation loss of the model is optimized by WGAN-GP(Wasserstein Generative Adversarial Nets-Gradient Penalties)theory,and the network model is generated by pre-training with public data sets.Finally,the predictive protection of murals is realized by super-resolution reconstruction of murals.(2)The mural classification model of separable network.With Google Net model as the basic framework,the background features of murals are extracted shallowly with small convolution kernel,and then the larger convolution kernel cross is separated into smaller convolution kernels to extract the important deep feature information of murals.Soft thresholding is used to activate the scaling strategy to increase the stability of network training,and small batch random gradient descent algorithm is used to update the parameters.In the newly designed adaptive separation convolution module,complex features of murals are extracted,and murals are finally classified according to different styles and types of murals.This thesis has completed the construction of the ancient mural data set,the research on the method of super-resolution reconstruction of the mural,and the research on the classification method of the different painting styles of the mural.Use crawler technology and data enhancement methods to collect the mural data set,and then use super-resolution reconstruction technology to optimize the resolution of the murals,and finally classify the murals,breaking through the existing technical bottleneck of the digital protection of ancient murals.The obstacles encountered in exploring the research value of protecting murals provide a new technical solution,and propose a stable enhanced generative adversarial network model and an adaptive separation convolution model,expand the deep learning algorithm,and provide a reference for the stable training of other network models.