Research And Application Of Thangka Image Inpainting Based On Deep Learning

With the support of a large number of image data sets,the image inpainting model based on deep learning can better capture the advanced features and deep structure hidden in the image,and achieve an inpainting effect closer to the image characteristics and human vision.Combining the image inpainting technology based on deep learning to simulate and inpaint the irregular damaged area of the Thangka image can not only ensure the efficiency and quality of the inpainting,but also avoid secondary damage to the Thangka.It has important social and cultural significance and academic practical value for realizing the digital protection of Thangka.Based on the digital protection of Thangka,this thesis realizes the research and application of deep learning in Thangka image inpainting.The main work of this thesis is as follows:(1)At present,there are few authoritative and publicly available mask dataset for the irregular damage characteristics of Thangka and the normalized Thangka image dataset for deep learning.Based on this situation,this thesis proposes to construct a Thangka image irregular mask data set that focuses on the damage characteristics of Thangka,mainly linear and circular,and supplemented by arcs.The clear and intact images in the Thangka resource library are manually preprocessed one by one to obtain1666 normalized Thangka image data sets suitable for deep learning applications.(2)Aiming at the limitations of traditional methods to inpaint irregular damage,based on the irregular mask data set of Thangka,this thesis proposes to use a partial convolution-based deep learning model for inpainting training.For the Thangka image data set,a fine-tuning strategy is proposed.The model is improved and trained by freezing the batch normalization layers of the coding network,so that the total loss function of the model on the Thangka image data set is reduced by about 0.25 compared with the initial stage of fine-tuning.Taking peak signal-to-noise ratio,structural similarity,fidelity of visual information and inpaint time as evaluation indicators,a large number of experimental comparisons and analyses are carried out with TV model,Criminisi algorithm and Patch Match algorithm.The results show that with the support of the Thangka image data set and the irregular mask data set for the damage characteristics of the Thangka,the deep learning model with transfer learning solves the problems of insufficient ability of traditional methods to deal with irregular damage of Thangka and slow inpaint speed,and the inpaint results are more accurate and more in line with the characteristics of Thangka images.(3)Aiming at the obvious repair traces that may appear in the weak texture area of the Thangka image inpainted by the model,this thesis proposes a combined inpaint method.On the basis of the model inpainting results,the Patch Match algorithm is used to perform partial secondary inpainting of the inpainting traces to improve the final inpainting effect of the Thangka image.After experimental comparison,the combined inpaint program has also obtained ideal experimental results.(4)On the basis of a series of work,starting from the function and performance of the Thangka image inpainting system,the requirement analysis and function analysis of the Thangka image inpainting system have been completed.The designed and implemented Web-based Thangka image inpainting system can meet the requirements of Thangka image inpainting efficiency and quality.