自然图像上色研究综述

图像上色是指从灰度图像中恢复图像的色彩信息，一张灰度图像可以有多个合理的上色结果，具有多模态的不确定性.另外，在上色过程中经常会出现颜色溢出、颜色暗淡等问题.传统的上色方法耗时长且效果不佳.最近，深度学习技术的应用使图像上色取得了显著的进展.文章将自然灰度图像上色分为4类：基于涂鸦的图像上色、基于参考图像的图像上色、全自动图像上色和基于文本的图像上色，并对这4类自然图像上色的技术方法进行回顾与总结；然后，讨论分析了深度学习给上色带来的影响、目前使用的损失函数以及评价指标；最后，总结了图像上色中存在的问题和未来的研究发向，为后续图像上色的研究提供参考.

A survey of the research on natural image colorization

Image colorization refers to recovering the color information of an image from a grayscale image. A grayscale image can have multiple reasonable colorization results with multimodal uncertainty. In addition, the problem of color bleeding and dull color often occurs during the colorization. Traditional colorization methods are time-consuming and ineffective. Recently, the application of deep learning techniques has made significant progress in image colorization. This paper classifies natural grayscale image colorization into four categories: scribble-based image colorization, reference-based image colorization, fully automatic image colorization, and text-based image colorization. This paper reviews and summarizes the technical methods of natural image colorization from these four categories, and discusses and analyzes the impact of deep learning on colorization, the loss function and evaluation indicators currently used. Finally, the limitations of current image colorization and future research directions are summarized, which provides references for subsequent image colorization research.