【作者】 刘克飞；

【导师】 魏颖； 穆世杰；

【作者基本信息】 东北大学 ， 计算机技术， 2009， 硕士

【摘要】 近年来,场景监控在各个领域得到了广泛的应用,但由于天气、光照、摄像机运动、信号传输、成像系统自身限制等原因,给图像的清晰程度带来了影响。图像质量不可避免的降低,轻者表现为图像不干净,难于看清细节；重者表现为图像模糊不清,连概貌也看不出来,大大降低了监控系统的可用性。因此,序列图像的清晰化处理算法研究对场景监控系统具有重要意义。在查阅大量资料和进行多种方法实验的基础上,本文针对序列图像相邻帧中物体重复出现且变化不大及信息可相互弥补的特点,将图像配准方法与超分辨率重建算法相结合,提出了一种针对全局运动的序列图像清晰化处理新方法。通过对序列图像进行处理,可生成较原图像更为清晰的高分辨率图像。本文从视频监控系统影响图像质量的主要因素入手,分析了成像模型,对序列图像超分辨率重建涉及到的问题进行了深入探讨。针对全局运动的序列图像,引入互信息度作为图像配准的相似测度,结合Powell优化算法的一般思想,讨论了基于两者的图像配准方法。在此基础上,紧接着分析研究了POCS(凸集投影)超分辨率重建算法,并针对凸集投影算法收敛慢、去噪能力弱等特点,在估计初始值方面作了改进,采用中值综合的融合方法综合图像序列相关信息来重新估计初始估计帧,从而加强了图像清晰化和去噪效果。最后运用Matlab编写实验软件平台,进行算法验证,提高了研究效率,并分别对标准图像和真实监控图像进行实验,通过实验结果分析,证明了本文算法可行有效。更多还原

【Abstract】 In recent years, video scene monitoring system has been widely used in various fields. However, due to some reasons such as weather, lighting, camera movement, signal transmission and restrictions of imaging system itself, the clarity of the images has been decreased, and the quality of images has been inevitably reduced seriously. Images may be not clean and difficult to be seen in the details; More seriously, images are not clear, and they even can’t be recognized in general. All these defects decrease the functions of video scene monitoring system greatly. So research of restoration algorithms from sequence images is important for video scene monitoring system.Based on a lot of information and a variety of experimental methods, this thesis has combined the image registration method and the super-resolution reconstruction algorithm in accordance with the characteristics that the objects in adjacent frames turned up repeatedly, changed little and made up for each other in the sequential images, and provided a new method to clear the image sequence against the global campaign. Through processing the sequence images that are not clean and have noise, it can generate a better high-resolution image than the original images.This thesis has analyzed the imaging model and the issues that refer to the reconstruction of the image sequence deeply, starting from the main factors that video scene monitoring system affect the image quality in. Against the image sequence of the global campaign, it has introducted the mutual information as the measure of the similar degree of the image registration, combined with the general ideas of Powell optimization, and discussed the registration method based on the two ones. On this basis, it has analysed the super-resolution reconstruction algorithms of the POCS (convex projection), improved the estimation of the initial value against the characteristics that the convex projection algorithm has the slow convergence and the weak capacity to eliminate noise, and adopted an approach of median integration to estimate the initial estimate frame that has related information of the image sequence, and then enhanced the image clarity and the effects of eliminating noise. Finally, the software platform that is programmed by Matlab to verify the algorithm has improved the efficiency of research and tested the standard images and the real monitor images respectively. Through the analysis of the results, the method is proved feasible and effective.更多还原