Research On Ancient Mural Completion Technology Based On Two-Stage Generative Adversarial Network

The ancient Chinese murals represented by the Mogao Grottoes in Dunhuang have a long history and high cultural value.However,with the change of time,the ancient frescoes were inevitably damaged by factors such as wind,sand,moisture or human theft.Therefore,timely restoration of the discovered damaged murals can preserve their economic and cultural value to the greatest extent.Traditional manual restoration not only has higher technical requirements for restoration experts,but also has the possibility of damage to murals once the improper operation.Therefore,the digital restoration of the damaged mural can not only avoid the damage caused by human error to the mural itself,but also make the complete digital image of the mural easy to preserve and spread.Image completion algorithm is the key technology of digital image restoration.It can automatically complete the missing area by inputting the position information of the damaged image and the damaged area to generate a complete image.In recent years,depth generation model has become the mainstream method of general image completion because of its powerful feature representation and reasoning ability.Existing general image completion algorithms based on depth generation model mainly have the following problems :(1)the region generated by the algorithm is sometimes blurred or affected by artifacts;(2)The boundary of the defect area is fuzzy,inconsistent with the context structure and so on.These factors will affect the look and feel of the final completed mural.This paper will explore how to use the general image completion algorithm to solve the problem of mural restoration and improve the effect of mural restoration through the improvement of the algorithm level.Specifically,the main contributions of this paper are as follows:(1)Proposed to image completion algorithm based on boundary connection on the mural completion task,during the whole process of fresco fill image of high frequency and low frequency information,decoupling,the edge completion stage to get the edge of the imaginary in the effect of prior information promotion mural completion stage,and the common large face data sets and the Dunhuang frescoes completion competition training data set and verify the effectiveness of the proposed algorithm.(2)In judging the image in order to solve the completion algorithm considering the local image consistency shortage problem,the introduction of the contain a combination of local discriminant and global discriminant criterion method to improve,to the completion of the whole image via the global discriminant criterion,the global consistency of the image,at the same time to make up for the whole area as the center of local image using local consistency judging device for judging,The consideration of local features of input image is strengthened,and the stability of confrontation training is improved.(3)To effectively use context information and consider the context in the generated image consistency to generate completion as a result,in the image completion module generator network module,introduces the semantic consistency attention in the process of generating image block,joined for completion and image known regional and generated correlation consideration,improve the quality of image generation.In this paper,the performance of the baseline completion algorithm and the improved algorithm are evaluated by the evaluation indexes such as structural similarity and peak signal-to-noise ratio.Meanwhile,the results are tested and trained on the Celeba dataset commonly used in image completion tasks and the Dunhuang Fresco Completion Competition dataset.Qualitative and quantitative analysis shows that the authenticity and quality of images completed by the model proposed in this paper have been improved,which verifies the effectiveness of the method proposed in this paper...