【作者】 闫华；

【导师】 刘琚；

【作者基本信息】 山东大学 ， 通信与信息系统， 2007， 博士

【摘要】 在医学诊断、卫星成像和视频监控等应用中，我们经常需要高分辨率图像提供更多的细节和信息。由于CCD和CMOS传感器的分辨率级不够高，采集的图像在放大时经常容易产生失真。如果我们单纯通过改进硬件系统的性能来提高分辨率级，在技术方面很难实现突破，同时也会带来难以承受的成本增加，会严重阻碍应用的推广，所以通过软件增强采集图像的分辨率，突破硬件系统的限制具有非常大的研究意义。超分辨率技术是一种图像融合技术，它能够利用信号处理的方式从同一场景的多个观测低分辨率图像中获得高分辨率图像或序列，从而克服传感器和光学制造技术的限制。目前，这一分辨率增强技术已经成为最活跃的研究领域之一。本论文从低分辨率成像的数学模型出发，回顾了基于重构和基于学习的超分辨率算法，并进行系统分析和研究，最后提出了一些解决超分辨率问题的方案。主要成果概括如下：1．提出了基于边缘投影的图像配准方案。本论文利用Canny算子能同时照顾图像的强、弱边缘，具有较好的抗噪性能，较高的边缘定位精度和较低的误检率等特点，将它作为边缘提取算子来提取参考图像和待配准图像的边缘。然后利用Radon变换将提取的边缘在不同的射线方向上进行投影，来准确地反映边缘走向，采用边缘投影的最大互相关准则进行图像的粗配准，获得图像间的旋转角和像素级的平移量。最后进一步利用数据拟合的方法对平移估计进行细化，配准的精度由细化的重采样间隔确定；2．提出了考虑图像配准误差的规整化超分辨率重构算法。a．由于图像的配准并不能完全获得与原始运动一样的参数，在超分辨率问题中，本论文将位移帧差引入的误差称之为配准误差，用来反映交互通道的配准误差信息，并将其作为噪声均值与低分辨率观测通道内的加性高斯噪声相结合得到新的高斯噪声。b．本论文基于Miller规整的思想逐像素限定了两种约束，并通过规整化参数将这两种约束相结合，构造规整化代价函数。规整化参数根据观测通道内的加性高斯噪声污染的程度和通道间配准误差逐像素进行选择，自适应地控制相对于观测低分辨率图像的逼真度和期望的高分辨率图像的先验平滑之间的折中。c．本论文提出了两种递归实现规整化算法的方案：同步递归和并行递归方案重构高分辨率图像。d．本论文在消除光学模糊的同时提出对有限传感器阵列的平滑和理想采样分别处理来消除传感器的平滑。图像的上采样操作利用最邻插值代替补零插值来抑制边缘波纹。3．提出了基于误差-参数分析的超分辨率盲辨识方案。a．在单通道参数辨识中，本论文通过误差-参数分析获得单通道的误差-参数曲线，根据曲线的变化趋势来经验地选择表征模糊函数的参数；b．在多通道单参数辨识中，由于通道间重构误差的交互影响，本论文将误差-参数分析与搜索算法相结合，根据最小重构误差准则获得需辨识的参数。搜索算法的结合有效降低了计算量，实现模糊参数的快速辨识；4．提出了基于独立分量分析的超分辨率人脸假想和识别方案。将高分辨率人脸图像作为混迭信号，利用FastICA算法对大量高分辨率人脸样本进行学习获得独立的人脸特征，即高分辨率基图像。将高分辨率基图像进行下采样获得近似独立的低分辨率基图像。将待识别的低分辨率人脸图像在低分辨率基图像空间投影，利用无约束最小二乘获得投影坐标，保持投影坐标不变，将其作为相应的高分辨率基图像的混迭系数，最后进行高分辨率人脸的假想和识别。综上所述，本文针对基于重构的超分辨率问题的三个方面：图像配准、将病态的超分辨率问题转化为良态的规整化及盲模糊辨识三个方面进行了研究；将独立分量分析运用到基于学习的超分辨率中，实现了超分辨率人脸假想与识别。论文最后总结了该领域急待解决的问题和下一步的研究重点。更多还原

【Abstract】 In most electronic imaging applications, images with high resolution (HR) are desired and often required. A HR image can offer more details that may be critical in various applications, such as medial diagnosis, satellite imaging, and video surveillance so on. However, the current resolution level of CCD and CMOS image sensors can not make a captured image have no visible artifacts when it is magnified. It is expensive and difficult to increase the current resolution level by improving hardware performance, which is not considered effective.Super-resolution technique, a kind of image fusion, can use signal processing techniques to obtain an HR image (or sequence) from observed multiple low-resolution (LR) images of the same scene, and overcome these limitations of the sensors and optics manufacturing technology. Recently, such a resolution enhancement approach has been one of the most active research areasThis dissertation introduces mathematic model of LR imaging, reviews reconstruction-based and learning-based SR algorithms, and carries on analysis and research. Finally some significant schemes are proposed. In general, the main ideas are included as follows:1. Edge-projection based image registration scheme is proposed. In this scheme, firstly, Canny operator is used to extract the edges of reference image and sensed image, since it can detect strong and poor edges at the same time, and represents good performance, such as anti-noise, high precision of edge orientation and low ratio of error detection. Then, Radon transform can exactly reflect edge orientation, and has two useful properties: shift-invariant and rotation-invariant, so it is used to project the extracted edges along different angles. Next, cross-correlation-based approach is used to estimate global rotation and shift. Finally, data fit is used to improve the accuracy of image registration. Registration precision is dependent on fine resample distance. 2. Regularization SR reconstruction considering inaccurate image registration is proposed.a. Since the performance of the registration algorithms cannot ensure estimated motion parameters coincident with original ones in certain environments, displaced frame differential (DFD) caused by an inaccurate registration should be considered in the reconstruction procedure. The error caused by the DFD, which is called cross-channel registration error, is considered as mean value and combined with observed within-channel AWGN to construct a new Gaussian noise..b. Since registration error is pixel-wise different, two kinds of constraints are regulated according to the idea of Miller’s regularization. Regularization parameters connect the two constraints to construct regularized cost function. In term of contamination degree of within-channel AWGN to LR image and cross-channel registration error, regularization parameters are adaptively selected to balance the fidelity of desired HR image to LR images and its smoothness.c. Two kinds of recursion implementation schemes are proposed to reconstruct a desired HR image, that is, synchronous recursion and parallel recursion.d. The smoothness and ideal sampling operation of limited sensor array to HR image are respectively implemented. In reconstruction procedure, the smoothness is removed as well as optical blur. Up-sampling operation is implemented by nearest interpolation to restrain edge artifact.3. Error-parameter analysis based SR blind identification is proposed. Blur operator can be represented by parameter model, so the identification of blur operator is converted to parameter identification.a. In single channel parameter identification, error-parameter curve is obtained by error-parameter analysis, and blur parameter is chosen experimentally in term of curve tendency.b. In multi-channel single-parameter identification, error-parameter analysis is combined with searching method due to alternate influence of reconstruction error between multiple channels. Minimal reconstruction error rule is used to identify blur parameter. The combination with searching method reduces computational cost and makes speedy parameter identification come true.4. Independent Component Analysis (ICA) based SR face hallucination and recognition is proposed. In the scheme, training HR face samples compose mixed data matrix. FastICA algorithm is used to separate the mixed data for obtaining independent HR source face images. Then the HR source face image is down-sampled to obtain approximately independent LR source face images. An input LR face image is projected in the space formed by the LR source face images, and unconstrained least square method is used to obtain projection coefficients. The projection coefficients are retained and linearly mix the corresponding HR source face images. Finally HR face hallucination and recognition are performed.In sum, three aspects about reconstruction based SR algorithm, such as image registration, regularization which makes ill-posed SR problem well-posed, and blur identification, are investigated in this dissertation, and ICA is used in learning based SR algorithm to realize HR face hallucination and recognition. Finally, the problem to be solved related to this research field and future research topics are summarized, furthermore, the prospect of the developing tendency is analyzed as well.更多还原