【作者】 杨文波；

【导师】 朱明；

【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) ， 机械电子工程， 2014， 博士

【摘要】 随着航空技术日新月异的发展，航空成像技术已经在地形测绘、土地和森林资源调查、城市建设、铁路和公路建设以及军事侦察等诸多领域得到了广泛的应用。但在航空摄影的过程中，图像的质量受诸多因素的影响。由于受太阳光散射引起的空中雾的亮度、航空摄影中飞机的震动、曝光时间内飞机运动引起的像点移位、成像系统物镜的残余像差、感光材料的性能以及飞行姿态等因素影响，使实际获得图像的分辨率较低，图像质量较差。因此，研究航空图像的超分辨率重构，从图像处理的角度提高航空图像的分辨率，具有重要的应用意义。本文着重研究航空图像超分辨率重构的一些关键技术问题，研究成果如下：(1)单帧图像超分辨率重构。同一场景不具备拍摄多帧图像的条件，针对这一问题，本文重点研究了线性插值、3次Ｂ样条插值、O-MOMS插值和Keys插值四种多项式插值重构算法，给出了四种线性空间移不变的多项式插值公式，但是线性空间移不变插值技术易导致图像边缘平滑，为了解决这一问题，本文提出了一种线性空间移变自适应插值算法，该算法是对有偏距离插值算法的改进，对多项式插值中的像素值赋以不同权值，像素值的权值因子大小取决于其邻域像素值不一致性，权值因子能使图像局部均匀一致，有助于保持图像局部连续性，提高重构图像的质量。(2)序列图像超分辨率重构。本文将互信息配准技术应用到航空图像序列的配准中，用来提高图像配准的精度与鲁棒性。针对航空图像大数据量的特点，本文提出了一种基于先验信息的超分辨率重构算法。该算法把图像复原和信息融合分开处理来提高运算效率，利用比较简单的非迭代算法进行图像复原，再用小波变换的方法将图像序列中所包含的冗余信息和互补信息融合到一帧图像中，最后用多项式插值重构高分辨率图像。该算法的核心是去模糊，由于模糊矩阵是分块Toeplitz循环阵，因此存在非奇异矩阵使其对角化，避免了对大型稀疏矩阵直接求逆，加快了运算速度。本文分别给出了线性最小均方误差法、最大信息熵法、正则化方法三种去模糊的非迭代数值算法。针对航空图像模糊核有时难以进行有效的估计，导致观测模型无法精确的建立，本文提出了一种基于最大公因子法的盲超分辨率重构算法。该算法把图像复原和信息融合分开处理，是以模糊核的盲估计为基础的。对于图像的盲复原，本文采用二维最大公因子盲复原算法，由于二维最大公因子盲复原算法仅针对两帧图像进行恢复而非整个图像序列，因此本文通过序列图像的每一帧信息，重新生成一帧集合了所有图像中的信息的观测图像，给出了序列图像的二维最大公因子盲复原算法。(3)亚像元成像增加探测器的时间和空间采样频率，可以提高亚像元成像系统空间分辨率。但是探测器采集到的数据有混叠，重构得到的高分辨率图像发生模糊，分辨率远未达到理想值，为了解决这一问题。本文提出一种针对多线阵亚像元成像的超分辨率重构算法。首先，在高分辨率网格上建立插值模型；然后，辨识插值重构图像在线阵列方向和扫描方向的模糊核，得到整幅图像的模糊核；最后，采用带有Neumman边界条件的梯度平滑正则化模型去除模糊。更多还原

【Abstract】 With the rapid development of aviation technology, aviation imaging technology has beenwidely applied in such fields as topographic mapping, land and forest resources survey, urbanconstruction, railway and highway construction, military reconnaissance, and so on. But in aerialphotography, the image quality, which is influenced by many factors such as intensity of fogcaused by sunlight scattering, vibration of aircraft, image motion in exposure, aberration ofobjective lens in imaging system, performance of photographic materials, flight attitude and so on,is inferior. Therefore, it is of great significance in application to study super-resolutionreconstruction of aviation images aiming at the improvement of its resolution from the perspectiveof image processing. The paper puts emphasis on some of the crucial technical issues onsuper-resolution reconstruction of aviation images, the research results of which are as follows:(1) Single-frame image super-Resolution reconstructionThe capture of multi-frame images at the same scene is not available, thus, the paper doesresearch on the four algorithms for polynomial interpolation: linear interpolation, cubic B-splineinterpolation, O-MOMS interpolation and Keys interpolation, and presents four linearspace-invariant formula of polynomial interpolation. However, linear space-invariant interpolationtends to smooth edges. In order to solve the problem, the paper proposes adaptive linearspace-variant interpolation, which is an improvement of warped-distance interpolation. The pixelsin polynomial interpolation are assigned with different weights, and the weights depend on theasymmetry in neighboring pixels. The homogeneity in images yielded by weights contributes tothe improvement of the quality of reconstructed images.(2) Sequence image super-resolution reconstructionMutual information registration is applied to registration of aviation images to improve theaccuracy and robustness of image registration.In consideration of a huge mass of data in aviation images, a super-resolution reconstructionalgorithm based on a priori information is presented in the paper. The algorithm breaks thesolution into image restoration and fusion consecutively to improve efficiency. It usesnon-iterative algorithm to restore images, and then fuses redundant information andcomplementary information in images into one-frame by wavelet transform, and last reconstructs a high-resolution image by polynomial interpolation. The essence of the algorithm is deblurring.The degradation matrix is the one of block Toeplitz-To-block circulant, thus there is a non-singularmatrix to achieve diagonalization property, avoiding direct inversion of a large sparse matrix andimproving the computational efficiency. The three non-iterative algorithms for deblurring, whichare linear minimum mean square error, the maximum entropy, and regularization methodrespectively, are presented in the paper.In view that the estimation of blur kernel in aviation images is not always available, whichleads to failure to establish a precise observation model, the reconstruction algorithm of blindsuper-resolution based on two-dimensional greatest common divisor(2D-GCD) is proposed in thepaper. The algorithm, based on blind estimation of blur kernel, breaks the solution into imagerestoration and fusion consecutively. As to the blind restoration of images, the algorithm of blindrestoration by2D-GCD is employed in the paper. The2D-GCD algorithm is only to benefit fromtwo observations rather than all observations. Therefore, a new observation image incorporatingthe information from all the observations is generated, and The2D-GCD algorithm of sequenceimage is presented in the paper.(3) Spatial resolution of the sub-pixel imaging system can be improved by an increase oftemporal and spatial sampling frequency in detectors. However, there exists aliasing in datacollected by detectors and blurring in an image with high-resolution, as a result of which theresolution is far away from the ideal value. To resolve the problem, an algorithm ofsuper-resolution reconstruction concerning multi-linear array sub-pixel imaging is put forward inthe paper. First of all, an interpolation model on high-resolution grid is established. Next, blurkernels in an image with high-resolution are identified in linear array and scanning directionrespectively, from which the blur kernel in a frame is obtained. Last, a model of gradientsmoothing regularization with Neumman boundary conditions is employed to deblur.更多还原