Research Of Mural Image Restoration Technology Based On Multi-Stage Deep Generative Model

In order to obtain better mural restoration results,this paper takes the intelligent virtual restoration of Tibetan murals as the starting point and conducts the research on deep learning-based mural image restoration technology.The main content of this paper is as follows:(1)For the problem of repairing structural defects in murals,this paper proposes a Multi-stage Image Damage Inpainting Network Model(MIDI),which can be used in most scenarios where image structure is missing.This network model consists of two sub-models: Inpainting Submodel and Optimizing Submodel.Inpainting Submodel can gradually repair the image defect area with the help of Global Adaptive Attention Mechanism and partial convolution char-acteristics.Optimizing Submodel uses the Multi-scale Convolution Residual Blocks to optimize the detailed texture of the generated image features by Inpainting Submodel.In addition,this paper conducted comparative experiments and corresponding ablation experiments for the MIDI model on three standard public datasets.The experimental results proved its excellent repair performance and the effectiveness of each module of the model.Finally,this paper brought MIDI into the repair experiment of real mural image structure missing,confirming its better performance in this field.(2)For the problem of restoring non-structural defects in murals,this paper proposes a Multi-stage Image Degradation Recovery Network Model(MIDR).By synthesizing the artifiMural art is one of the important cultural heritages of the Chinesenation,showing the culture of people that has been passed down for thousands of years.However,after a long time of natural erosion,many have shown varying degrees of damage.With the advancement of computer technology,digital image processing technology has developed rapidly.By deeply mining image feature information,convolutional neural networks have shown extraordinary strength in the field of image restoration.Therefore,combining deep learning im-age restoration technology to digitally restore Tibetan murals can make up for the low efficiency and secondary damage of manual restoration and has great social significance for mural pro-tection.However,most of the commonly used deep learning image restoration methods show certain limitations in the field of mural restoration: for example,in the repair tasks of structural defects such as mural image shedding and attachments,most methods have poor repair effects when facing large-area irregular and rich-textured defect areas? In the restoration tasks of mixed non-structural defects such as mural image decolorization and noise,some algorithms use artificially synthesized mural degradation images for training,and some can only restore a single degradation factor.Applying them to naturally formed real mural degradation images may cause serious performance degradation.cial unstructured defect image corresponding to the real intact image,it transforms network tasks into learning mapping relationships between them and transfer tasks between synthetic images and real non-structural missing images.This paper uses a Full Dimensional Variational Autoencoder Submodel Based on Generative Adversarial Networks to improve transfer learn-ing effects and a Transformer-based Feature Mapping Submodel to enhance mapping learning capabilities.Finally,this paper compared MIDR with some popular algorithms on the collected real mural dataset.The results showed its better non-structural recovery ability.At the same time,corresponding ablation experiments also proved the effectiveness of each module.