Research On Old Photo Restoration And Super-resolution Technology Based On Deep Learning

Old photos are the main carrier for recording the elapsed time and the old photo restoration has important research and application value.With the passage of time,old photos may be severely damaged during the long-term storage.The degradation factors include the scratches,the creases,the blurring,the color-fading and so on.In addition,due to the limitation of the resolution of the cameras,the information recorded by the old photos cannot meet the current demands.Therefore,the restoration and super-resolution technologies of old photos have become a hot topic.The core issue of this research is how to restore the complex hybrid defects existed in those old photos while improving the image resolution.We conduct the relevant research to tackle the issue and propose the efficient restoration and super-resolution algorithms of old photos based on deep learning.The main contents and innovations of the thesis include: 1)We study and model lots of degradation factors that exist in those old photos.We establish two dataset for experiments,namely the synthetic degradation dataset(SD-SET)and the real old photo dataset(ROP-SET).2)We propose a class-attribute guided old photo restoration algorithm,which completes the restoration of the global defects by extracting the latent class-attributes of old photos.The class-attributes are also used to guide the local detail enhancement module to improve the details of old photos.In addition,we propose the dual discriminators to further improve the restoration quality of old photos.3)We propose an old photo restoration algorithm that combines the reference priors and the generative priors.We use the reference images to guide the restoration processes of old photos.At the same time,a pre-trained generative model is used to reconstruct the missing information of old photos.This algorithm achieves the good restoration quality and generalization performance.4)We propose an attention-enhanced old photo super-resolution algorithm,which completes the super-resolution reconstruction through the multi-scale content-aware attention module and the orientation-aware attention module.Our algorithm can reconstruct better details and structures of old photos than other algorithms.We use the public old photo dataset(POP-SET),the self-establish datasets(SD-SET and ROP-SET)and the super-resolution datasets(DIV2K,Set5,Set14,Urban100,BSD100 and Manga109)to complete the training and the test of all algorithms.The results show that our algorithms are significantly better than other mainstream algorithms both the subjective quality and the objective metrics.Moreover,our algorithms have more powerful capabilities of the old photo restoration than the mainstream commercial software.