基于多尺度注意力半监督学习的老照片划痕修复

老照片由于长时间的磨损或保存不当，会出现照片的划痕损伤。随着深度学习在图像重建中的应用，基于深度学习方法能够在纹理修复的基础上获取图像的语义信息并预测语义内容，使老照片修复的整体效果更加符合客观事实，但利用深度学习进行老照片划痕修复缺乏学习所需数据集。提出一种基于半监督学习的老照片划痕自动修复的方法，创建划痕合成数据集SynOld用于网络训练，同时搜集真实的划痕老照片用于训练和测试，将合成数据集和真实老照片加入网络学习，两者共享网络参数，并通过鉴别器来区分网络生成图像与真实图像。对于合成数据集有监督的分支采用均方差损失、感知损失和对抗损失约束训练，对于真实老照片无监督的分支采用总变差损失控制训练。实验结果表明，相比于多尺度特征注意力网络的监督学习方法，该方法在合成数据集SynOld和真实老照片上都具有较好的修复效果。

Scratch Repairing of Old Photos Based on Multi-Scale Attention Semi-Supervised Learning

Old photos acquire scratches and other damages due to long-term wear or improper preservation.The application of image reconstruction, based on the deep learning method, can repair the texture of the image using semantic information and can predict the semantic content.This ensures that the reconstructed photos are accurate as compared to the original photo, however, one of the difficulties for learning-based scratch repair is the lack of data sets. An automatic scratch repair method for old photos is proposed based on semi-supervised learning, whereby a synthetic dataset SynOld is created for network training and additional real old photos were collected for training and testing.Both synthetic data sets and real old photos were utilized for network learning.There are shared network parameters, and a discriminator is added to distinguish whether an image is generated by the network or is a real image.For the supervised branch, the mean square error loss, perceptual loss, and adversarial loss are used.For the unsupervised branch, total variation loss is used to achieve the purpose of enhanced reparation of the old photos.Experimental results demonstrate that compared with the supervised learning method of multi-scale feature attention network, the proposed method has a better repair effect on both the synthetic dataset SynOld and the real old photos.