Abstract:

Image super-resolution (SR) is an important type of image processing technology for improving image and video resolution in computer vision. In recent years, thanks to the success of neural networks, image super-resolution technology based on deep learning is booming, which is undoubtedly the mainstream direction of super-resolution technology research. In this paper, the research progress of image super-resolution is summarized. First of all, the existing super-resolution techniques are summarized and introduced according to their input and output. What??s more, the single-image super-resolution technique based on deep learning is discussed from two aspects of supervised and unsupervised learning respectively, and some representative latest super-resolution reconstruction technologies are introduced. Then, the related problems of super-resolution technology, namely performance evalua-tion indicators and standard data sets, are discussed, and several typical algorithms are compared by experiments. Finally, the future research trend of image super-resolution algorithm is prospected.

Key words: deep learning, image super-resolution, supervised learning, unsupervised learning

摘要：

图像超分辨率（SR）是计算机视觉中提高图像和视频分辨率的一类重要技术。近年来，得益于神经网络的成功，基于深度学习的图像超分辨率技术正在蓬勃发展，这无疑是超分辨率技术研究的主流方向。对超分辨率工作进行综述。首先，总结目前已有的超分辨率技术，根据其输入输出进行分类介绍；其次，将基于深度学习的单图像超分辨率技术分为有监督学习和无监督学习两类进行论述，并对部分具有代表性的最新超分辨率重建技术进行总结分类介绍；然后，讨论了超分辨率技术的相关问题，即性能评价指标、标准数据集，进而对几种典型算法进行实验对比；最后，对图像超分辨率算法未来的研究趋势进行展望。

关键词: 深度学习, 图像超分辨率, 有监督学习, 无监督学习