The Research On Color Restoration Method Of Mural Image Based On Multivariate Feature Optimization

In the course of more than 5,000 years of history,the Chinese civilization has preserved many colorful cultural heritage.Among them,the Dunhuang Grotto murals represented by the Mogao Grottoes are unique.With diverse themes,huge scale and rich content,the Dunhuang murals vividly reflect the social life,religious beliefs,artistic aesthetics,political and military affairs,inner aspirations and other information in ancient China and the Western regions from the 4th to 14 th centuries with unique painting techniques and artistic presentation methods.They have important scientific,social,historical and artistic research value.For a long time,due to the influence of time,climate change,human destruction and other factors,the cave murals have generally produced a variety of diseases,which lead to the mural defects,scratches,coating,fading,discoloration and other problems,which seriously affect the development of research and work in related fields of mural.In recent years,digital image processing technology has made a lot of progress in the fields of mural protection,display and restoration.Therefore,it is very necessary to carry out the research of digital mural image protection and restoration by relying on digital Dunhuang resources and image processing technology.In this paper,color transfer and style transfer techniques based on deep learning theory are used to study the color restoration of faded/discolored mural images.The main research work is as follows:1.In order to solve the problems that it is difficult to retain the detail texture of the mural image and the degradation of small area defects and scratches in the process of mural image color restoration,a method of mural image color restoration based on the Transfer and Refine network was proposed.Based on encoder-decoder network framework,a two-stage network of the Transfer and Refine is constructed to process mural images with different resolution modes.In the low resolution mode of mural image,the Transfer network is used to restore the color of faded/discolored mural image,and the attention mechanism is used to reduce the phenomenon of detail loss which is easy to produce in the low resolution image,so as to preserve the detail texture of mural image to the maximum extent.Then,in the mural image high resolution mode,the Refine network is used,and the mural image edge is combined to further optimize and restore the mural image detail texture,and improve the color restoration effect.Experimental analysis shows that this method can accurately restore the color of faded/discolored mural images and has a better objective evaluation score.2.In order to solve the problem that it is difficult to select reference mural images similar to faded/discolored mural image in conventional mural image restoration method in mural color restoration task,and thus affect the quality of color restoration,a mural image color restoration method based on double reference optimization was proposed.In this method,two reference mural images are used to restore the faded/discolored mural images.Image optimization module is used to suppress multiple degradation such as noise which is common in faded/discolored mural images.Multi-scale features of mural images are extracted by encoder-decoder network coding,and a feature fusion module is constructed to optimize image features.The semantic confidence of reference mural image and faded/discolored mural image is calculated by the double reference guidance module to realize similarity matching between image regions and style fusion of two reference mural images.Finally,the color restoration of faded/discolored mural images is realized by using fusion features.Experimental analysis shows that this method is effective in restoring the faded/discolored area of mural image,and is better in subjective visual experience and objective evaluation results.