Study Of Mural Image Restoration And Super-Resolution Reconstruction Based On Generative Adversarial Network

Ancient murals have great research value as precious cultural heritage in China.However,due to the impact of human activities,nature and other objective conditions,the existing frescoes have many different types of diseases and are in urgent need of protection.The resolutions of the captured mural images are easily limited by equipment and other factors,making it difficult to meet the needs of digital conservation of murals.At a Time of growing development of deep learning,it brings a wide range of technical support to fresco image restoration and super-resolution reconstruction.To do so,this paper develops a study of large-area damage restoration and super-resolution reconstruction of mural images according to generative adversarial networks,specific research efforts are summarized as follows::(1)To tackle the problem that there is no standardized mural image dataset,this paper collects mural images from various channels and uses data enhancement algorithms after preprocessing to complete the construction of mural image dataset.(2)For the problem of difficult feature abstraction and inconsistent context structure when restoring large damaged mural images,a double-discriminative generative adversarial network approach to mural image restoration has been proposed.Where the generator is improved on U-Net with the addition of null convolution to expand the perceptual field and fully extract the available information of the image;using a dual discriminatory network combining local and global to ensure overall consistency of the restored image;in particular,WGAN-GP is used to enhance network training stability.On the basis of the test comparison with the classical restoration algorithm on the constructed mural image dataset,the restored fresco images by this method achieved better results in terms of detailed texture and overall structure.(3)In order to address the problems of low resolution of mural images and blurred texture details,a super-resolution reconstruction method of GAN mural images fused with residual attention is proposed to improve.The model deleted the batch normalization layer in the network that affected the quality of the reconstruction.The use of high frequency features is enhanced by adding a residual attention module combining residual structure and channel attention mechanism to the generative network.And optimized the discriminative network structure to improve its discriminative performance.A joint loss function combining content loss,perceptual loss and adversarial loss is also used to further optimize the training results.After verification,the proposed super-resolution reconstruction algorithm achieves a great reconstruction outcome in the mural images,and the reconstructed mural images are more realistic and clear.