Abstract: This paper comprehensively surveys the development of example learning-based single image Super-Resolution（SR） technique. The example learning-based single SR methods apply machine learning technology to estimate the high-frequency details missing in a given Low-Resolution（LR） images to produce a High-Resolution（HR） one with sharper edges and finer textures by learning the mapping relationship between the LR and HR images from a training set. According to how to use the examples and how to learn the relationships in the process of SR reconstruction, the existing example learning-based SR methods can be divided into five typical categories, i.e., k-Nearest Neighbors（k-NN）learning-based methods, manifolds learning-based methods, dictionary learning-based methods, example multi-linear regression-based methods, and deep learning-based methods. The main idea of each type and the representative SR approaches are introduced in detail. Then the SR results of six representative SR methods are compared and analyzed. Finally, the deep insight into example learning-based SR methods and promising future directions are summarized.

Key words: example learning, manifolds learning, example regression, dictionary learning, single image super-resolution, image quality assessment

摘要： 全面综述了基于学习的单帧图像超分辨重建技术的研究与发展。基于学习的单帧图像超分辨重建借助机器学习技术，通过学习低分辨与高分辨图像之间的映射关系估计低分辨图像中丢失的高频细节，以获得边缘清晰、纹理细节丰富的高质量图像。根据超分辨重建过程中实例样本使用方式和学习算法的不同，已有基于学习的超分辨重建方法可分为五种类型，包括基于[k]近邻学习的方法、基于流形学习的方法、基于字典学习的方法、基于实例多线性回归的方法和基于深度学习的方法。对每类方法的主要思想和具有代表性的方法进行了详细介绍，对六种具有代表性的基于学习的超分辨重建方法的重建结果进行了比较和分析。最后，对基于学习的超分辨重建技术的未来发展趋势进行了展望。

关键词: 实例学习, 流形学习, 实例回归, 字典学习, 单帧图像超分辨, 图像质量评价