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

Dunhuang grotto murals are treasures of Buddhist art in China and the world.These murals contain ancient Chinese cultural customs,religious beliefs,production and life,natural environment and other content,with high artistic and research value.For a long time,Dunhuang murals have been degraded to varying degrees due to various disease factors caused by objective conditions such as environment and climate.However,due to hardware equipment or external interference in the process of collection,transmission and display of degraded murals,the collected mural images have low resolution,texture blur,detail defect and other phenomena.Image super-resolution reconstruction refers to the reconstruction of clear high-resolution images based on blurred low-resolution images.The use of super-resolution reconstruction technology can significantly improve the display effect of mural images.The current super-resolution reconstruction methods still have problems such as difficulty in accurately generating highfrequency texture information and serious edge artifacts when reconstructing mural images containing a large amount of complex texture information.In this thesis,the super-resolution reconstruction method of mural image is studied according to the composition structure and content characteristics of mural image.The main research work is as follows:(1)Aiming at the problem that the existing super-resolution reconstruction methods extract image features on a single scale,and it is difficult to reconstruct the high-frequency structural information of murals by using the feature dependencies of different scales,a super-resolution reconstruction method of mural images based on pyramid attention is proposed.This method uses the self-similarity of the mural image content texture to construct the mural feature pyramid and obtain the multi-scale features of the mural image.The attention is used to calculate the correlation of similar textures in different spatial locations of multi-scale features,so as to capture more image high-frequency information to reconstruct the mural texture details.At the same time,the hole residual dense block is constructed to obtain the global features of the image,and the semantic consistency of the local region of the mural image is constrained to eliminate the edge artifacts.The experimental results show that compared with the benchmark method ESRGAN,the proposed method has better objective evaluation index and subjective visual effect,and improves the reconstruction effect of mural image detail information.(2)The mural contains a large amount of texture information with a certain artistic style,and the existing super-resolution reconstruction method is difficult to use the low-quality mural prior information to accurately generate the mural texture.A super-resolution reconstruction method of mural image based on similarity reference is proposed.This method uses the repaired high-resolution mural image as the reference image and constructs the mural image pyramid,and reconstructs by establishing the correlation between the reference image and the low-quality mural image at different scales.In the feature extraction stage,the convolutional layer and self-tuning pixel attention are used to extract the residual features of the reference image.In the texture transfer part,Transformer transfer reference image texture features are introduced to enrich the high-frequency texture of low-resolution images.At the same time,spatial adaptive normalization is used to eliminate the distribution difference of color and brightness between reference image and low resolution image.In order to improve the applicability of the method,a residual feature distillation block is constructed between each stage to reduce feature redundancy.The experimental results show that the proposed method has achieved good results in terms of objective evaluation indicators and subjective visual effects,while reducing model parameters and has good applicability.