Research On Old Image Restoration Based On Deep Learning

Old images record people’s past clothing,food,housing,and transportation,record the changes of the times,and also record the great transformation of China in the past century.Restoring these old images obtained from historical photography and film archives is crucial for preserving,displaying,and disseminating past cultural heritage to future generations.Traditional old image restoration mainly uses physical methods,digital methods,or image processing methods to manually repair frames by frames,which requires a large amount of time and manpower,and has a high cost.In recent years,with the development of deep learning,the restoration of old images based on deep learning has become a research hotspot,and has achieved good results beyond traditional methods in many scenarios.However,due to the complex reasons for the quality degradation of old images,existing methods still have problems such as poor performance,insufficient naturalness,and insufficient model lightweighting in certain scenarios.This paper mainly focuses on the denoising and super-resolution research of old images,focusing on issues such as significant structural degradation noise removal and excessive sharpening of super-resolution processing results.The main work and contributions of the paper are as follows:(1)For the old image denoising task,an attention mechanism based old image denoising algorithm is proposed,which can effectively solve various mixed degradation problems in old images,and the repair results for large structural degradation such as vertical stripes and scratches are more realistic and natural,and can perfectly match the surrounding pixels.This algorithm adopts a novel Mixed Feature Attention(MFA)module proposed in this paper on the network structure.This module integrates channel attention,spatial attention,and pixel attention,which can effectively improve the repair of structural degradation.It creatively integrates the mixed feature attention module into the triple domain transformation network(reference[23]),improving the generalization ability of old image denoising models.In order to verify the effectiveness of the algorithm,the paper constructed a dataset consisting of thousands of real and synthetic images.The test results on the dataset showed that the proposed method outperformed other comparative methods in both objective indicators and subjective effects.(2)For the task of super-resolution reconstruction of old images,a Transformer based super-resolution algorithm for old images is proposed,which makes the results of super-resolution reconstruction more realistic and natural without excessive sharpening.In terms of network structure,this paper proposes a new transformer foundation block,Channel Transformer Block,which combines channel attention.It improves the image detail processing effect by increasing the range of receptive field.At the same time,this paper introduces CBAM(Convolutional Block Attention Module)in the super-resolution reconstruction network to improve the network’s expression ability.To verify the effectiveness and rationality of the algorithm,this paper collected thousands of high-definition old images combined into a new dataset based on the DIV2K+Flickr2K dataset.The experimental results show that the old image super-resolution algorithm proposed in this paper can achieve higher quantitative indicators and better visual effects.(3)In order to verify the effectiveness and rationality of the old image denoising and super-resolution reconstruction algorithms proposed in this paper,this paper designs and implements a B/S architecture old image restoration system.Users can access the system through a browser to facilitate and quickly complete the restoration of old images.The system client is developed using the Vue front-end framework and Antd(Ant Design Vue)component library,supporting users to preview uploaded images and download the repaired results.The server uses the Express framework and middleware development to handle various requests from the client,and calls the Pytorch network model to repair old images.The old image restoration system designed in this paper is lightweight and convenient,with low coupling between functional modules,and a good user experience,which has high practical value.