The Implementation Of Old Photo Restoration Based On Deep Generation Model

Old photos retain valuable historical image information,but today’s existing old photos are often damaged to varying degrees.Although these old photos can be digitally processed and then restored,the restoration of old photos involves multiple fields of image restoration and various degradation types.At present,there is no unified model for the restoration of various degradation types of old photos,so there is still a great space for development of the restoration technology of old photos.In this paper,the depth generation model is used to deeply study the image super-resolution reconstruction and image damage filling technology involved in old photos restoration.The main work contents are as follows:(1)Aiming at the fuzzy defects of old photos,this paper introduces the image super resolution technology into the fuzzy restoration of old photos,and generates Super Resolution Generative Adversarial Networks in image super resolution reconstruction.The self-attention layer is introduced,and the Charbonnier loss replaces the L2 loss as the content loss.In addition,the calculation of perceived loss utilizes the eigenvalues of the Deep Convolutional Neural Network(VGG19)proposed by the Visual Geometry Group of Oxford University before activation.The validation test of the model is carried out in SET5 and SET14 datasets,and compared with the benchmark model,Both the Peak Signal to Noise Ratio(PSNR)and Structural Similarity(SSIM)of the experimental indicators have been improved,verifying that the improved model is better than the original network.Finally,the model is generalized to the collected old photo data to verify the experimental results.It can be seen that the human eye perceived the sharpness of old photos to a small extent.(2)For defects such as scratches and spots in old photos,this paper first made data sets and trained scratch detection network.Then,in order to narrow the gap between synthetic old photos and real old photos,a kind of Variational auto-encoder based on spatial mapping transformation method was proposed.Variational Auto-Encoder(VAE)and Generative Adversarial Networks(GAN)combined photo restoration network model.In this model,the generated old degraded photos and the real old degraded photos were mapped to the same feature space by a variational autoencoder,and the mapping relationship between the generated old degraded photos and the original pictures was learned to realize the restoration of old photos.Then defect detection and facial enhancement network were introduced to increase the visual performance of the restored photos.Finally the experimental results on the test data set and pix2pix network and image super-resolution reconstruction based on the attention mechanism against comparing network,though the experiment index value less volatile,but the actual visual experience has a certain promotion,at the end of the paper will be the result of the experiment I was with you and old photos repair found when comparing two commercial APP,The algorithm in this paper is more obvious to the complex background damage perception,and has a certain color enhancement effect.