Research On Ancient Painting Image Inpainting Based On Partial Convolutional Neural Network

In China,Shaanxi in particular has abundant ancient painting resources.These ancient paintings have very important artistic,scientific,historical and archeological values,but the existing conditions of these paintings are very bad,and most of them are facing serious problems about aging,disease,damage or extinction.Ancient paintings that have been damaged have seriously affected the appreciation and inheritance of digital cultural heritage represented by paintings.In order to improve the digital quality of ancient paintings,this thesis will use deep learning theories and methods to study corresponding technical measures to reconstruct high-quality ancient painting images for damaged low-quality ancient painting images.Due to the complex structure of ancient painting images and the small number of samples,it is difficult for deep learning algorithms that perform well on natural image inpainting tasks to be effective for ancient painting image inpainting tasks.How to build an ancient painting image inpainting network with good restoration effects and strong migration capabilities has become the focus and difficulty of the research in the field of image inpainting.The research goal of this thesis is to build a network model of ancient painting image inpainting with high restoration performance.In this thesis,two space-varying activation functions with better performance are proposed by combining Re LU and Leaky Re LU activation functions;By improving on the basis of partial convolution,a dual-domain partial convolution module is proposed;Improve the existing irregular hole image inpainting algorithm model based on partial convolution,and propose a new ancient painting image inpainting algorithm model.The main research content and achievements of this thesis are as follows:(1)The activation functions used by the current mainstream image inpainting network are only the element product between the input and the weight mapping,only the value of a single element in the input is considered,and the influence of surrounding elements is ignored.Regarding the defect,this thesis proposes space-varying activation functions that fully consider the spatial position relationship between adjacent elements to correct multi-scale feature mapping;Moreover,the irregular hole image inpainting network based on partial convolution is selected as the basic network,combined with the space-varying activation function and the leaky space-varying activation function to improve the basic network,and one image inpainting networks with higher repair performance are proposed.And conduct experiments on two public data sets.The experimental results show that the image inpainting network after adding the space-varying activation function module and the leaky space-varying activation function module has significantly improved repair performance compared with the basic network.(2)Aiming at the existing mainstream image inpainting network,the convolution module only extracts the spatial domain features of the pixels when extracting image features,ignoring the frequency domain feature information of the pixels,resulting in the extracted feature information not being detailed and comprehensive enough,which causes the network model to be more ancient repair results of holes in the complex structure of painting images are ill.In response to this defect,a dual-domain partial convolution module was first proposed based on the partial convolution module;Secondly,a network model of ancient painting image inpainting based on dual-domain partial convolution is constructed,and experiments are compared with the most advanced image inpainting algorithms on the Places365-Standard data set and ancient painting data set.The results show that the ancient painting image inpainting algorithm based on dual-domain partial convolution is superior to the most advanced image inpainting algorithms in subjective vision and objective evaluation indicators.