Restoration Of Old Photos Related To Railway Based On Deep Learning

The old railway photos are the epitome of the history of Chinese railway development in the past century.They not only show the historical trajectory of Chinese railway development,but also provide precious historical materials and important basis for the historians to study the history of railways,and have very important practical significance and historical value.However,due to factors such as the age,immature technology at that time,or improper preservation,the old railway-related photos preserved today have problems such as yellowing,fading,mottled,dark and old,and poor resolution.In this thesis,the image super-resolution reconstruction technology and image inpainting technology are applied to the field of railwayrelated old photo restoration,and the variational autoencoder network and generative adversarial network are used to study the old photo restoration technology,so as to realize superresolution reconstruction of old railway-related photos and repair of scratched areas.The main contents of this thesis are as follows:1.Construct a dataset for restoration of old railway-related photos.The old railway photo dataset consists of 100 old color photos and 100 old black and white photos.The resolution of the image is about 500*375,including along the railway,old railway stations,trains various scenes such as locomotives and railway workers,the images are all from Baidu images,Google images and Beijing Railway Museum.The scratch mask dataset contains 500 scratch mask images.The labelme annotation tool is used to manually mark the scratches on the damaged old photos to make mask fold scratch images,and the mask dataset is augmented by image processing methods such as rotation and symmetry transformation.2.Super-resolution reconstruction of old railway photos based on LPIPS-GAN.The LPIPSGAN network is extended on the model proposed by CIPLab,adding an attention module between the residual dense module and the upsampling module of the generator,and at the same time improving the activation function in the Dense Block module,choosing to use the Mish activation function instead of the original Re LU activation function to further improve the performance of the model.The experimental results show that the PSNR value and SSIM value of the LPIPS-GAN model reach 21.247 and 0.551,respectively,which are 0.271 and 0.022 higher than the baseline model,respectively.The LPIPS value of this model reaches 0.203,which is 0.05 lower than the baseline model.3.Scratch repair of old railway photos based on LATENT-VAE.The LATENT-VAE network draws on part of the network structure in the model proposed by Microsoft Research Asia,and increases the number of residual modules in the mapping network.The mapping network uses eight residual modules and a global information extraction module containing a non-local module to further Improve the quality of scratch repair for old railway photos.The experimental results show that the PSNR and SSIM values of the LATENT-VAE model reach20.785 and 0.668;the LPIPS value reaches 0.105,which are 0.119 and 0.019 lower than the two old photo restoration software,respectively.