【作者】 刘刚；

【导师】 戴明；

【作者基本信息】 中国科学院研究生院（长春光学精密机械与物理研究所） ， 光学工程， 2012， 博士

【摘要】 随着信息时代的快速发展和图像处理技术的日益普及，科学研究和实际应用对高分辨率图像和图像序列的分辨率的要求越来越高。高分辨率图像能提供更多细节信息，这对于图像的分析和处理起着重要作用。然而,在某些应用场合中,由于光学物理器件、处理器性能、信道传输带宽以及存储容量的限制，常常获得的图像的分辨率较低，并且有时候无法或是难以通过直接更改硬件配置方式突破这些限制。因此，在现有条件限制下来提高图像和视频序列的分辨率，这一直是人们研究的热点。多帧图像的超分辨率重建技术已被证明是一种解决上述问题非常有效的方法。它利用同一场景的多幅图像间互补的信息，采用信号处理的方法进行融合得到一幅分辨率增强的图像。这一技术在不改变硬件条件的情况下，可以有效地提高图像的空间分辨率，可以为目前在光学物理器件性能受限的情况下获取高质量的图像，因此具有广泛的应用前景，对该技术的研究具有重要的理论和实用意义。本文对图像和视频序列的超分辨率重建技术进行理论分析，对运动估计和基于正则化的超分辨率重建等若干关键问题进行了研究，同时也对单帧图像分辨率增强进行研究。主要创新工作和研究成果如下：1)对图像配准参数的Cramer-Rao下边界进行推导。对基于梯度方法的平移图像配准进行分析，推导出其配准偏差，并利用分析结果构建新的配准方法以提高配准精度。这种基于迭代配准算法在输入为高斯噪声影响下是最优的。提出两种针对欠采样视频序列的配准算法来提高配准精度。一种是捆绑调节方法，通过加强运动流一致性来减小配准方差。这种新颖的鲁棒权值能够对奇异帧进行检测和对奇异运动进行抑制。另一种是采用基于高分辨率‘融合’图像的配准方法。从多帧LR图像重建的HR图像含有的噪声更少并且混叠程度也更少。这样，可以更可靠地计算出强度和梯度信息。因此可以提高配准精度。2)在一些应用中，需要对单帧图像进行分辨率增强。提出一种基于抗混叠contourlet变换的分辨率增强方法。改进contourlet变换能够对图像提供一个很好的稀疏表达，这非常适合保留轮廓和边缘。所提方法能够有效提高重建图像的边缘一致性。首先认为给定的低分辨率图像是原始高分辨率图像进行小波变换的的低通输出。利用简单小波变换插值得到高分辨率图像的初始估计。然后利用观测约束和稀疏约束迭代交替进行加强。3)分析了当前绝大多数超分辨率算法的限制，为了克服上述限制,提出一种混合范数的超分辨盲复原算法，其中的权值包括全局权值和局部权值。提出一个广义的权值自适应的混合L1和L2范数的代价函数。权值会根据配准误差和噪声分布自适应得改变并且惩罚图像中配准错误的部分。本章算法对奇异值具有很强的鲁棒性，同时超分辨率图像和模糊算子可以联合地估计出来。主观评价和客观评价都表明了本算法的有效性。4)提出一个基于正则化代价方程的图像序列超分辨率方法，能够同时重建估计出高分辨率序列的所有帧。因为只在代价方程的先验项中加入了运动信息，因此与其它方法相比,本章所提方法具有很好地数据保真性和鲁棒性,能够增加对运动误差的控制进而提高算法的鲁棒性。利用收敛速度快的共轭梯度法进行代价函数最优化求解。与其它流行的图像序列超分辨率算法进行比较，所提出的算法重建质量更高并且计算代价更小。更多还原

【Abstract】 With the rapid improvements of video and image processing technologies inrecent years, the demand for high-quality video and image sequences grows fast. Ahigh-quality image always contains further detailed information of targets, and it is ofgreat value for analysis and post-process. But in some application areas, under limitedoptical elements, processors, channel bandwidths or storage capacities, the imageresolution is always unable to meet our needs. Furthermore, it is impossible or hard tobreak the limitations. So, how to enhance the spatial resolution of video and imagesequences under these limitations becomes a research hotspot.Super-resolution image reconstruction technique has been proved to be anefficient technique to solve the above problems. It fuses complementary informationof several low resolution images by signal processing methods to get a highresolution image. It can enhance the spatial resolution of images effectively withoutany upgrade of current equipments. This technique provides us with an efficientapproach to obtain high-quality videos and images subject to the constraints of opticaldevices, processors or communication channels. Thereby, it is worthy of notice bothfor academic studies and applications, and it permits widespread deployment.The dissertation investigates several key issues of super-resolution of image andimage sequences including registration estimation, and regularization based imagereconstruction, and has obtained many results.The main contributions and innovation points of the dissertation are as follows:1)The Cramer-Rao bound for a general parametric registration with specialattention to two cases: translational and2D projective registration is analyzed. Biasesof the gradient-based estimator are then derived. This dissertation corrects the shift bias in an iterative manner and shows that the iterative gradient-based estimator isoptimal. presents two techniques for improving registration of (under-sampled)image sequences. The first one is Bundle adjustment, for example, reduces thevariance of registration by enforcing consistent flows from SINGLE image to anothervia any of the intermediate image routes. Registration on a high-resolution\fusion"image is another technique that improves multi-frame registration.2)In some applications, it needs to enhance the resolution of a single frame. a novelimage interpolation algorithm that uses the new contourlet transform to improve theregularity of object boundaries in the generated images is proposed. Assumes thegiven low-resolution image is the lowpass subband of an wavelet transform of theunknown high-resolution image, while all the coefficients in the highpass subbandshave been discarded. Firstly, By using a simple wavelet-based linear interpolationscheme as our initial estimate. We then attempt to improve the quality ofinterpolation, particularly in regions containing edges and contours, by iterativelyenforcing the observation constraint as well as the sparseness constraint.Experimental results show that our new algorithm significantly outperforms linearinterpolation in subjective quality, and in most cases, in terms of PSNR as well.3)Some limitations of conventional super-resolution methods are analyzed. To solvethese problems, proposes a general cost function that consists of weighted L1-andL2-norms considering the SR noise model where the weights are generated from theerror of registration and penalize parts that are inaccurately registered. Both thesuper-resolved images and blurring operators are jointly estimated. The objective andsubjective results are shown to demonstrate the effectiveness of the proposedalgorithm.4)A new class of algorithms for super-resolution of image sequences is proposed.This class of algorithms estimates simultaneously all frames of a sequence byemploying an iterative minimization of a regularized cost function. Similarly to othersuper-resolution techniques, the proposed approach exploits the correlation amongthe frames of the sequence. This correlated information helps to improve theresolution of the captured images. By employing the motion information only in theprior term of the cost function, the proposed method achieves a better fidelity andmore robust performance. An implementation utilizing Conjugated Gradient, withfast convergence, is presented.The proposed method is compared with other classicalmethods in the literature and the experimental results clearly indicated that theproposed method produces images with higher quality and lower computationalcost. Besides, the proposed method, with Huber norm, is very robust to outliers andprovides edge-preservation.更多还原