基于深度学习与图像锐化策略的老照片修复方法

An old photo restoration method based on deep learning and image sharpening

年代久远的老照片是多种图像降质因素的复合,为解决当前研究仅针对某一降质影响,无法实际应用于真实场景下的老照片修复任务,且端到端的老照片修复输出对结果多样化不具有鲁棒性的问题,本文构建了一个基于深度学习的交互式老照片修复系统。针对图像褪色、画质模糊、图像缺损问题,实现对老照片清晰度的提升。根据老照片中人像占比分类修补破损区域,全图像着色与区域多样化着色,在引入极少量用户信息的情况下将输入老照片映射到完整的彩色输出图。实验结果表明,该修复系统可以获得较好地满足人眼视觉观察、多样化的修复结果。

Old photos with a long history have the combination of multiple image degradation factors.In order to solve the problems that many current studies only focus on one of image degradation factors which can not be applied to reallife scenes,and that the end-to-end repair process is not robust to the diversification of results,this paper proposes an interactive old photo repair system based on deep learning.The goal of system is to solve the problems of image fading,image blur and image defect to improve definition of old photos.According to the proportion of people’s heads in the whole old photos,the whole images at all areas are colored,and mapping the old input to the complete color output can be achieved only with a very small amount of user information.The experimental results show that the proposed system can obtain a variety of repair results that meet the needs of human visual observation.