Research On Dunhuang Mural Inpainting Algorithm Based On Deep Learning

In the long history of splendid culture,Dunhuang murals depict the beliefs and customs of various ethnic groups,reflect the literature and art of the society at that time,carry important historical and cultural information,record the rise and fall of Chinese civilization,and have an extremely important position in the history of human development.However,due to factors such as climate,environment,and human operations,there are various diseases such as nail peeling,cracks,and so on in the murals in the grottoes.If manual inpainting is used,it requires high inpainting personnel,requires a long inpainting cycle,and is irreversible,resulting in a risk of secondary damage to the murals,which seriously hinders the inpainting work.Therefore,the introduction of digital technology to inpainting Dunhuang murals can effectively reduce the complexity and cycle of inpainting,while reducing the damage to the mural body.This paper delves into the current state of domestic and international research on image inpainting,analyzes its advantages and disadvantages,and optimizes and improves on the original deep learning model.The specific content is as follows:1.Aiming at the problem that it is easy to ignore the difference between damaged feature information and complete feature information in the process of mural images inpainting,which leads to unreasonable inpainting,a mural inpainting model fusing dynamic feature selection and pixel level channel attention is proposed.Firstly,a U-Net based network generator is designed to encode and decode damaged mural images;Secondly,an effective transferable convolution module is used to achieve flexible extraction of effective feature information by dynamically selecting the sampling space location.Then,a region synthesis normalization module is used to reduce the expected and variance deviation between the inpainting region and the complete region,thereby enhancing the selection and utilization of effective feature information;Finally,a pixel level channel attention module is designed at the decoding layer to enhance the weight of effective features while enabling the model to learn effective features from distant spatial locations.Experimental results on a Dunhuang mural dataset show that the algorithm can use effective information to inpainting irregular damaged mural images with varying proportions of mask regions,and is superior to the comparison algorithm in terms of subjective visual perception and objective evaluation indicators.2.In the process of inpainting mural images,the range of learning areas required to inpainting different types of missing mural images is different.Most existing inpainting networks based on large receptive fields often involve more unreasonable features that affect the inpainting effect.A local inpainting network and global inpainting network are combined to model and propose a local and global refinement mural image inpainting network.Firstly,a inpainting network with large receptive fields is used to learn the overall structure of the image and complete rough inpainting of damaged images;Then,a local refinement network with small receptive fields based on channel attention is designed to learn and refine the surrounding feature information of the damaged area.The texture information of the damaged edge on the coarse inpainting result is refined,and channel attention is added to further weaken the impact of remote invalid features on the inpainting result;Finally,a global refinement network with a large receptive field based on a half instance normalization block is designed to globally optimize the locally refined inpainting results.Context attention enables the network to learn effective features with a high degree of matching from a remote feature space.Adding a half instance normalization block ensures that the inpainting area has a similar distribution to the normal area and further performs feature enhancement.Simulation results show that compared to traditional inpainting models,this method can achieve large area damaged mural inpainting,and the inpainting effect meets the global structural and semantic consistency.