Research On Super-resolution Reconstruction Of Ancient Mural Images Based On Deep Learning

In the process of ancient mural image acquisition,due to the constraints of imaging equipment and environmental factors,there are distortion phenomena such as low resolution,unclear texture and fuzzy details in the captured images,which affect the viewing effect of mural images.Therefore,as a software solution,image super resolution can reconstruct high resolution images from low resolution images,which has become a hot topic in the field of image processing and computer vision.However,the existing image super-resolution technology mainly deals with natural images,and its application in the field of ancient murals is still in its infancy.Unlike natural images,ancient mural images present rich lines and a large number of smooth colored areas.Based on this,this article using the migration of deep learning,learning,and other advanced technology,in view of the ancient murals in details based on the enhanced image super-resolution research,first on natural images training network model to study the mapping relationship between low resolution and high resolution image,and in the ancient murals fine network model in the image,so as to decrease the difficulty of network training model To improve the super-resolution reconstruction effect of ancient mural images.The main work of this paper includes:(1)The existing super-resolution reconstruction network ignores the inherent multi-frequency features of the image,the low-frequency information and high-frequency information are treated uniformly,and the context information extracted by the single-scale network model is limited,and there are many spatial redundancy in the features.To solve this problem,using the idea of octave convolution for reference,this thesis designs a super-resolution algorithm based on multi-frequency hierarchical residual network.The algorithm groups the extracted shallow features according to high frequency and low frequency through the feature separation module to reduce the spatial redundancy of feature tensor;Then,the multi-frequency feature mapping module is responsible for the internal feature updating of high and low frequencies and the information exchange between frequencies,so as to improve the learning ability of the network from different frequencies;Finally,the high-low frequency feature fusion module is responsible for the fusion of high and low frequency features to unify the spatial resolution of high and low frequency features,so as to facilitate the final reconstruction of the network.The experimental results show that the improved algorithm shows good results in the objective evaluation index and visual perception on the natural image datasets,which verifies the effectiveness of the method.(2)The current super-resolution reconstruction algorithm based on CNN extracts the local features of the image layer by layer through convolution.The network does not fully extract the global information,so it is difficult to recover the high-frequency information by using the global structure of the image.To solve this problem,in the work of super-resolution algorithm of multi frequency hierarchical residual network,a super-resolution network based on double-layer enhancement module is proposed.Firstly,an efficient transformer module is cited to adapt to the task of super-resolution reconstruction.It is used as the basic unit for global feature extraction in the low-frequency feature tensor group,so that the low-frequency features can better recover the lost high-frequency information from the global feature,and make full use of the advantages of CNN local perception and the ability of transformer global modeling.Secondly,the high-frequency feature mapping branch with CNN as the backbone and the low-frequency feature mapping branch with transformer as the backbone not only carry out in depth feature extraction with in the branch,but also exchange local information and global information between the branches.The fusion process greatly improves the global modeling ability and local modeling ability of the network.Finally,this method shows a good reconstruction effect on multiple natural image datasets.Through transfer learning,the reconstruction effect on ancient mural datasets is also better than several other advanced reconstruction methods.