A Deep-learning Research And Application On Thangka Inpainting Based On Params And Computational Methods

Regong art is an intangible cultural heritage in my country and the world,and Thangka is an important form of expression with distinct regional characteristics and cultural significance.Therefore,the digital protection and inheritance of Thangka is very important and urgent for it.However,due to the limited preservation conditions,many physical thangkas have problems such as fading,stains,cracks,creases,etc.,which have brought serious restrictions on their exhibition and inheritance.At present,convolutional neural network is still the main structure used by popular image inpainting algorithms,and it is mainly used in large-scale natural image datasets,and there are few studies on Thangka image datasets.The problem of image inpainting in thangka is studied.Since the current deep learning-based image inpainting methods are strongly data-driven and parameter-driven,they often perform poorly on data with less data and stronger rules.This paper proposes a lightweight image inpainting model LRGAN based on parameters and operations,and conducts a large number of comparative experiments.The main work of this paper is as follows:（1）In this paper,we construct an image inpainting model LRGAN based on parameters and operations.By combining the depthwise separable convolution method with the attention mechanism,the quality of the inpainted image is better when the parameters are greatly reduced;in addition,by using skip connections in the image inpainting network,the information in the early stages of the network is reused,which not only enhances the stability of the training stage,but also further improves the quality of the inpainted images.（2）This paper trains the LRGAN model on the public datasets Celeb A-HQ and Places2,and conducts multiple control experiments for image restoration at the same time,including CA（Contextual Attention）algorithm,PEN-Net algorithm,PConv algorithm,Gated Conv algorithm,which is related to this problem.The experimental results show that the LRGAN method achieves a balance between the amount of parameter calculation and the quality of repair,and the experimental results are better than other methods.Besides,To address the Thangka inpainting problem,this paper constructs a Thangka image dataset with 1568 images.And based on this database,the LRGAN method,the CA（Contextual Attention）algorithm,the PEN-Net algorithm,the PConv algorithm,and the Gated Conv algorithm are used to conduct comparative experiments.The experimental results show that the LRGAN method has a good effect in Thangka image restoration,and can achieve the same effect.In addition to the control experiments,in order to further verify the performance of the LRGAN model,this paper also conducts ablation experiments,which demonstrate the effectiveness of the mechanism proposed in this paper.（3）On the basis of the design algorithm,a thangka image restoration system based on the Flask Web framework is proposed,and the function design of the thangka image restoration system is completed,which can meet the requirements of the efficiency and quality of the thangka image restoration.