Research On Old Photo Restoration Algorithm Based On Deep Learning

Old photos record precious historical images and have extensive research value.However,in real life,there are a series of problems in the old photos.On the one hand,due to the limitations of the hardware and software of the shooting equipment at that time,old photos have problems such as low resolution and information defect.On the other hand,due to the poor preservation measures in the actual storage process,most of old photos have additional damage,blurring,scratches and other problems,and most of the image restoration algorithms cannot reasonably restore the old photos with the above complex conditions.In recent years,with the development of digital image processing technology and the need for the implementation of the digital of literature,the research in the field of old photo restoration is gradually carried out.In order to solve the above problems,based on deep learning technology,this thesis studies the defect restoration and fuzzy restoration of old photos.The main contents are as follows:(1)The current status of image restoration theories related to old photo restoration at home and abroad is summarized.The advantages and disadvantages of defect restoration models and super resolution models for scratch,stains,blur and so on,those are analyzed.The deep learning theories involved in the optimization algorithm for old photos restoration is expounded.(2)A scratch restoration model of old photos based on improved VAE-GAN is designed.Aiming at the structural defects such as scratches,creases and stains in old photos,the triple domains transformation model composed of Variational Auto-Encoder(VAE)and Generative Adversarial Networks(GAN)is used to restore the defect area.At first,the model maps the synthetic scratched old photos and the real scratched old photos into the same shared domain through VAE,and aligns the data domains corresponding to the two types of data,and then uses the mapping relationship between the synthetic scratched old photos learned by the model and the original photos to restore scratches and other defects.At the same time,the residual block in the original model is improved and replaced by the designed residual dense block with multi-channel attention.Finally,the scratch defect detection network and the facial enhancement network are combined to further improve the effect of the model on restoring structural defects.The experimental results indicates that compared with other ablation models based on VAE-GAN model and Cycle GAN,DIP,pix2 pix and other models,the VAE-GAN model with residual dense blocks with multi-channel attention can not only restore the structural defects of old photos,but also reconstruct the image with clearer texture and better visual effect,and further improve the PSNR value and SSIM value.(3)A fuzzy restoration model of old photos based on improved KXNet is constructed.For the restoration of fuzzy old photos,KXNet super-resolution model is introduced for high-definition processing.When old photos are reconstructed by the model,the fuzzy old photos are first input into the kernel estimation network,and the latest kernel is updated through operation.Then the updated kernel is input into the HR image estimation network for reconstructing clearer images.After several alternating updates,the best image is output,and then the restoration of the fuzzy old photos is realized.In order to further improve the fuzzy old photos,the multi-channel attention mechanism is introduced into the residual block part of the original model,which can capture more cross-channel features for convolution operations to reconstruct more reasonable local texture details.The experimental results indicates that compared with IKC,DANv1,DANv2,KXNet,KXNet+CBAM,and other models,the fuzzy restoration model of old photos based on improved KXNet not only improves the PSNR value and SSIM value at different scales,but also the texture details of the restored old photos are clearer than other models,and the visual perception of the old photos is improved.