Research On Restoration Method Of Damaged Mural Image Based On Deep Learning

The Dunhuang Caves are an artistic treasure trove filled with a vast array of mural elements.The mural paintings on the cave walls chronicle the long history of the Chinese nation from the 4th to the 14 th century,and they are rich in content and of great scale.The Dunhuang murals use brushstrokes to vividly depict information on the ancient ethnic customs,religious beliefs,artistic aesthetics,political culture,and more,making them of extremely significant research value in the areas of ethnicity,religion,art,and culture.For a long time,the Dunhuang Caves have been situated in a natural environment of harsh wind and sand conditions,compounded by the impact of human activities,resulting in various types of damage to the Dunhuang murals,such as detachment,cracks,scratches,blisters,inscriptions,and fading.In recent years,the image restoration technology has demonstrated practical value in the dissemination of cave murals,and it has become an indispensable tool in the digital protection of murals.Digital image restoration technology aims to preserve the background information of the image to be restored,and to fill in the holes in the unknown areas of the image to create a visually complete image.Existing image restoration technologies mainly focus on repairing damage to single content such as faces and street scenes.However,mural images are characterized by broad content and complex elements,which present difficulties in directly applying image restoration models for training.Therefore,it is necessary to study image restoration algorithms that take into account the artistic features of mural images.In order to improve the application of image restoration models in the restoration of damaged mural images,this thesis focuses on the feature extraction and image reconstruction methods of mural images in the research of deep learning-based image restoration technology.The specific contents are as follows:1.To address the difficulty of extracting artistic structural and textural features from mural images and applying them to the process of restoring damaged mural images,this paper proposes a method for restoring damaged mural images based on a structure-texture intersectional generative adversarial network.The proposed method introduces a standard attention module to reduce the weight of unimportant information and increase the weight and feature capture rate of important features,thus ensuring that the model obtains more texture and structural information to effectively restore more details of the damaged mural image.At the feature fusion stage of the network structure,a gate feature fusion module is used to exchange and combine the structural and texture information of the mural image,achieving an effective coupling of structural constraints and texture guidance.A deformable convolutional contextual feature aggregation module is designed to refine the generated content through adaptive learning and multi-scale feature aggregation.Experimental analysis shows that the proposed method for restoring damaged mural images based on a structure-texture generative adversarial network produces good restoration results for both artificially simulated and real damaged mural images.2.A method based on parallel cyclic feature reasoning is proposed for repairing damaged mural images,taking into account factors such as noise and local color fading that can affect the image.Traditional image repair algorithms for natural images struggle to obtain high-quality image features,which can result in significant artifacts and texture disorders in the repaired mural images.This proposed method constructs an improved parallel cyclic feature reasoning network structure,using multiple partial convolution layers for mask updating and obtaining the areas of the mural image that require repair during the region recognition stage.During the feature reasoning stage,a parallel encoder structure is designed to fully obtain mural image feature information.In the third layer of the encoder structure of the upper branch,a coherent semantic attention module is integrated to strengthen mural feature extraction,while a global-local attention module is introduced in the third layer of the encoder structure of the lower branch to enhance the correlation between the missing areas and background regions of the mural image.After recursive operations and utilizing the feature map generation characteristics of damaged areas during the image repair process,multiple different feature maps are generated and finally fused using an adaptive feature fusion scheme during the fusion stage to output the fully repaired mural image.Experimental results show that this method has significant subjective and objective improvements when applied to repairing damaged mural images with issues such as cracks,scratches,and detachment.