【作者】 杨芳；

【导师】 沈建冰；

【作者基本信息】 北京理工大学 ， 计算机科学与技术， 2011， 硕士

【摘要】 超分辨率图像技术的主要目的是由提供的一幅或多幅同一场景(信息相似但细节不同)的低分辨率(Low-Resolution,LR)图像中重建高分辨率(High-Resolution,HR)图像,它克服了图像获取过程中的限制病态条件,并能获得更好的图像内容,提高场景识别的准确性。该技术广泛应用于遥感识别、图像压缩、高清电视、安全监控、视频通讯、医疗诊断、资源探测等诸多领域,是当前图像处理和计算机视觉领域中的热门研究方向之一,具有很重要的理论研究价值。本文通过研究基于学习的图像超分辨率算法和基于梯度的图像超分辨率算法,以及图像细节增强相关方面的知识,提出了在能量优化框架下,结合边缘和细节的基于学习的单帧图像超分辨率算法。在本文中,首先简要介绍了图像超分辨率的概念、选题背景、研究意义及近十年的国内外发展情况。其次介绍了超分辨率技术的理论基础和超分辨率在空域上的各种算法,并简要分析了各种算法的优缺点。同时用Matlab工具实现了相关算法,得到放大后的图像。接着研究了利用马尔可夫网络来学习高分辨率图像和低分辨率图像之间的关系,并通过这种关系来进行超分辨率的基于实例的图像超分辨率算法和利用图像的梯度锐化度先验知识来指导超分辨率的基于梯度的图像超分辨率,进而通过实际编程实现上述两种算法,获得了相应实验结果。通过对实验结果的理论分析,认为可以结合边缘和细节两方面来提高经过超分辨率后的图像质量,最终提出了一种新的单帧图像超分辨率算法。本文重点介绍了能量优化框架下的基于学习的单帧图像超分辨率算法,章节3.3.1给出了超分辨率模型,并定义了能量方程和求解方法,章节3.3.2介绍了图像自身训练集的获得,章节3.3.3介绍了搜索方法,章节3.3.4给出了超分辨率算法的流程图及伪代码。从实验结果可以看出,本文提出的算法所得到的高分辨率图像在边缘和细节上都有所提高,具有更好的视觉效果。最后,对本文的研究内容进行了总结,超分辨率所面临的一些亟待解决的问题和下一步的研究重点,并对其未来的发展方向进行了展望。更多还原

【Abstract】 The main objective of super-resolution (SR) imaging is to reconstruct a higher-resolution image based on an image or a set of images, acquired from the same scene(similar information but different details) and denoted as‘low-resolution’images. It overcomes the limitation ill-posed conditions of the image acquisition process, can get a better image content and improves a scene recognition. This technology is widely used in remote sensing recognition, image compression, high-definition television, security monitoring, video communications, medical diagnostics, resource exploration and many other fields, is one of the most popular subjects in the current field of image processing and computer vision research, and has a very important theoretical research value.On the base of studies of the learning-based image super-resolution algorithms and the gradient-based image super-resolution algorithms, this paper proposes a learning-based single-frame image super-resolution algorithm combined with the information of edges and details under the energy optimization framework.Firstly, we briefly introduce the concept, the background of image super-resolution and then nearly a decade of significant domestic and international developments of it in this article. Secondly, we introduce the theoretical foundation of super-resolution technology, a variety of algorithms in the spatial domain, and a brief analysis of the advantages and disadvantages of these algorithms. At the same time we use Matlab tools to achieve these correlation algorithms, and gain the enlarged images. Then we study the example-based super-resolution through a relationship of the high-resolution images and the low-resolution images learned by the Markov network and the gradient-based super-resolution algorithm using the prior knowledge of image gradient, and then by the actual programming of these two algorithms to obtain the corresponding experimental results. Through theoretical analysis of experimental results that can be combined the information of edge and detail to improve the quality of the high-resolution image. Lastly, we propose a new single-frame image super-resolution algorithm.This paper focuses on the learning-based single-frame image super-resolution algorithm under the framework of energy optimization, section 3.3.1 gives a model of super-resolution, the definition of the energy equation and its solving method, section 3.3.2 introduces the own training set of images and section 3.3.3 the search methods, then section 3.3.4 shows the flow chart and pseudo code of the super-resolution algorithm. From the experimental results we can see that the obtained high-resolution images by the proposed algorithm are improved on the edge and detail, with a better visual. Finally, we summarize the content of this study, the problems to be solved that super-resolution has faced , the priorities of future research and prospect its future development direction.更多还原