基于深度学习的数字图像修复算法最新进展

Advances in Digital Image Inpainting Algorithms Based on Deep Learning

数字图像修复是一项利用计算机技术还原破损图像缺失信息实现自动图像修复的技术,其广泛应用于文物修复、电影特效制作、图像编辑等方面。早期数字图像修复算法可以对小区域缺失的图像进行近似真实的还原,而对大区域缺失图像的修复往往相似度低,内容模糊。近年来深度学习生成模型的进展为数字图像修复提供了新的技术手段,基于生成对抗网络和自编码器的图像修复技术成为研究热点。本文对数字图像修复算法最新进展整理归纳,按照算法类型将新算法分为三类,详细介绍每类算法的特点和不足;并在此基础上分别从卷积模式和网络结构优化两个方面详细阐述了最新研究进展和成果;最后对未来的研究方向进行展望。

Digital image inpainting is a technology of estimating the miss regions’information in defect images and restoring images as plausible as vision perception by computing algorithms.The applications of image inpainting are wild,such as cultural relics restoration,film particular effects generation,image edition.The pioneer digital image inpainting algorithm can work well for small missing regions,but when the missing regions are large,they usually failed to generate similar and realistic results.In recent years,the development of deep learning technology has provided new ideas for image inpainting.The generative models of deep learning have been a hot focus in digital image inpainting research,especially GANs and AEs.We summarize the algorithms based on deep learning technology and divides these algorithms into three types,then elaborately introduce the progress in both convolution methods and network structures,and finally discuss future research direction.