【作者】 蓝岚；

【导师】 黄东军；

【作者基本信息】 中南大学 ， 计算机应用技术， 2009， 硕士

【摘要】 人脸研究一直是计算机视觉、模式识别和计算机图形学领域中的热点研究问题之一。目前,通过监视器得到的人脸图像分辨率不高,以至于给人脸识别和跟踪等后续应用带来很大的困难。超分辨率图像重构(Super Resolution,SR)技术是一种基于信号处理技术来获得高分辨率图像的方法。SR技术的基本思想是,以若干模糊、有噪、频谱混叠的低分辨率(Low Resolution,LR)图像为输入,通过信号处理技术融合成一幅高分辨率(High Resolution,HR)图像。超分辨率图像重构技术在人脸或对象识别、遥感图像、视频监控、医学图像处理等领域都有着广泛的应用。本文主要研究单幅人脸图像的超分辨率重构技术,目标是提出一种更有效、实时性更好的算法来获得高分辨率图像。首先,本文全面回顾和评述了超分辨率图像重构技术的概念,基本方法和SR算法。在此基础上重点研究基于学习的图像SR算法。本文采用马尔可夫网络(MN)模型提出了一个新的框架描述重构机制。本文提出的算法采用对图像块搜索操作进行位置限制和检查图像分块间重叠区域水平兼容性的思想,降低了搜索的复杂度,提高了匹配相关性,加快了马尔可夫网络收敛,简化了隐层节点的计算。最后采用样本拼镶技术直接输出超分辨率图像。实验平台由VC++编程实现,实验中所用的人脸图像训练集采用24位灰度图像。实验结果证实本文提出的算法具有输出质量好、效率更高等特点,有一定的的实用价值。本文试图发展性能更好、更智能化的学习算法为以人脸图像为主的应用带来新机遇,并推动超分辨率技术自身的发展。更多还原

【Abstract】 Research efforts in face processing always is one of a hot spot research questions which in the computer vision, the pattern recognition and in computer graphics domain. Nowadays, the human face image resolution which obtains through the monitoring device is not high, the human face recognition and the track and so on the following application brings the very great difficulty for the human. The Supper Resolution (SR)technology is one kind method obtains the high resolution(HR)image based on the signal processing technology . The basic idea behind SR is the fusion of a sequence of low-resolution noisy blurred images to produce a higher-resolution image based on the signal processing technology. The SR in human face research, long-distance image remote sensing, video frequency monitoring, and medicine domains and so on has the very good application.We mainly study the single face image super resolution techniques with the goal aiming at an algorithm which is more simple, practical and suitable for real-time applications. First, this article comprehensively reviewed and narrated and commented the SR technology concept, essential method and SR algorithm. In this foundation key research based on study image SR algorithm. This article used Markov network (MN) model to propose a new Baye (MAP) frame description restructuring mechanism. We propose a novel algorithm that uses the location-restraint operation and uses the most compatible neighboring patches along horizontal dimension of the face to directly mosaic the high-resolution patches into the outcome. This can reduce the search order of complexity, enhance the match relevance, speed up Markov network restraining, and simplify the implicit strata node computation. The tests platform by the VC++ programming developmented, the human face image training sets which in the experiment uses 24 gradation images. The experimental result confirmed this article proposed the algorithm has the output quality well, an efficiency more higher characteristic, has certain practical value.更多还原