摘要图像超分辨率（SR）技术具有良好应用前景和研究价值，可以分为基于插值、基于重建、基于学习三类方法，基于学习的方法有很大的研究空间。本文介绍了几种常见的图像超分辨率方法，主要研究了一种基于两级框架的高斯过程回归（GPR）算法，由于高斯过程回归属于机器学习，和贝叶斯统计有一定联系，因此先引入了贝叶斯统计和机器学习的相关知识。研究了高斯过程回归的原理，协方差函数和超参数的选取，并将高斯过程回归应用到单个图像超分辨率。使用一个两级框架算法，根据输入训练集和观测值来训练高斯过程回归模型，得到最优超参数，实现超参数自适应，再根据测试点预测对应的预测值，从而实现了在没有额外训练集的情况下，仅由低分辨率图像得到高分辨率图像。实验结果显示，从视觉效果来看，两级框架高斯过程回归算法明显比一级高斯过程回归和双三次插值得到的高分辨率图像的质量好，边缘细节更丰富。最后，本文还介绍了高斯过程回归算法的改进方向。51315

关键词  超分辨率  高斯过程回归  协方差函数  超参数  机器学习

毕业设计说明书外文摘要

Title    Research of Image Super-Resolution Technology   

Abstract

Image super-resolution(SR) technology  has a good application prospect and research value.It can be pided into three kinds:based on interpolation, reconstruction and learning.In this paper,Several common image super-resolution methods are introduced,and a two level framework of Gaussian process regression (GPR) are studied.Due to Gaussian process regression belongs to machine learning and has relation with Bayesian statistics,the knowledge of Bayesian statistics and machine learning are introduced.The principle of Gauss process regression,covariance function and the selection of the Hyper-parameters are studied,and the Gauss process regression is applied to the single image super-resolution. Using a two stage framework, depending on the value of the input training set and observation to train Gaussian process regression model to obtain the optimal hyper-parameters, realize the self-adaption of hyper-parameters , then according to the test values to predict points.Without additional training set ,only use the low-resolution images to obtain high-resolution images. The experimental results show that:the two level framework of Gauss process regression algorithm is significantly better than the one level framework of Gauss process regression and bicubic interpolation.It’s can get high-resolution images with abundant edge details. At last, the improvement direction of Gauss process regression algorithm is introduced.

Keywords  super-resolution  Gaussian process regression  covariance function       hyperparameter  machine learning

目   次

1  引言 1

1.1  图像超分辨率技术的研究背景及意义 1

1.2  图像超分辨率技术的研究现状及研究难点 3

1.3  论文的研究内容 7

1.4  论文的行文安排 8

2  图像超分辨率的常见方法 10

2.1  基于插值的图像超分辨率方法 10

2.2  基于重建的图像超分辨率方法 11

2.3  基于学习的图像超分辨率方法 12

3  基于高斯过程回归的图像超分辨率实现方法 13