【作者】 曹聚亮；

【导师】 吕海宝；

【作者基本信息】 国防科学技术大学 ， 仪器科学与技术， 2004， 博士

【摘要】 图像超分辨率处理技术是指利用多帧关于同一场景的有相互位移的低分辨率降质图像来重建高分辨率高质量图像的技术。图像超分辨率处理技术可突破图像采集设备的分辨率限制，充分利用多帧图像之间的互补信息，实现像素级的图像信息融合。图像超分辨率处理和成像技术在遥感、军事、医学成像、公共安全等领域有很广泛的应用前景。它可以在一定程度上消除由于图像离散化和图像退化引起的空间分辨率下降的不利影响，弥补原有图像空间分辨率的不足，获得更加清晰的图像，可使输出图像的质量得到很大程度的提高，既改善图像的视觉效果，又便于计算机对图像进行处理、分析和理解。 本文围绕图像超分辨率处理和超分辨率成像技术研究了相关的频域和空域超分辨率处理算法、图像配准算法、微位移图像序列的获取方法、微变焦法、基于直方图修正的图像增强技术，以及在研究过程中提出的基于图像配准的二维微位移平台位置参数视觉测量方法等内容。具体而言，本文的主要研究内容和取得的成果包括以下几个方面： 1．分析并给出了图像超分辨率处理问题的频域和空域数学模型，给出了基于光流的分层迭代图像配准方法；提出条件行处理迭代法来实现频谱解混叠处理，该方法将正交投影和平行投影融合在一起，可有效提高处理速度。 2．从空域的角度给出了两种图像超分辨率处理算法：基于帧迭代的反向投影法和改进的快速行处理迭代算法。基于帧迭代的反向投影法通过建立虚拟全景图像及其迭代权矩阵，实现迭代反向投影的逐帧处理，从而可实时地对视频图像序列作超分辨率处理。在改进的快速行处理迭代算法中，首先建立了像素级的空域图像降质模型，采用四维非稀疏矩阵存储模型系数，避免了对大型稀疏矩阵的处理；进而对亚像素位移量进行量化处理，建立系数矩阵表，计算过程中通过实时查表的方式得到系数矩阵，使计算过程中对存储空间的需求由G级降低到K级；然后提出了改进的Cimmino迭代算法，算法充分考虑了系数矩阵的稀疏性，可大大加快迭代收敛速度。上述两种算法在处理中将超分辨率处理技术与图像镶嵌技术结合起来，不仅增加了图像的分辨率，也扩大了视场范围。 3．提出通过旋转双折射晶体来获得微位移图像序列的方法，并从理论上对该方法进行了分析和仿真；分析了微位移成像中位移量与超分辨率成像的关系及其对成像质量的影响，论证了常规微扫描法中等间隔扫描方案虽是最优的，但并非必需，只要适当增加观察图像帧数，不等间隔扫描也可以实现超分辨率成像，从而大大降低了对万国防科学技术大学研究生院学位论文微扫描机构的要求。 4.提出一种基于图像配准技术测量移动平台二维平移、旋转等参数的视觉测量方法。通过固定基准图和镜头，使CCD随平台在基准图的像平面内移动或旋转，再由图像配准技术对当前CCD接收到的图像在整个像平面内进行定位，从而可实现微位移平台位置参数的测量。初步实验结果表明系统的位移测量精度可达到亚微米级。 5.从空域的角度分析了微变焦法的原理。微变焦过程可等效为改变每个光敏元所对应物辐射率分布区域的位置和大小，由此建立了微变焦过程的空域模型及其简化模型。根据该模型可通过一个线性方程组来描述观察图像和目标高分辨率图像之间的关系，并通过求解该方程组得到目标图像。 6.研究了与图像超分辨率处理相关的基于灰度直方图变换的图像增强技术。为减小直方图均衡过程中图像的细节损失，首先提出两种改进的直方图修正方法:局部灰度修正法和灰度范围扩展法。另外，提出采用信息嫡的理论分析图像灰度分布与图像信息容量的关系，从而指导分析现有直方图修正方法的缺点，在此基础上又提出基于灰度资源分配原理的加权嫡直方图修正法。这三种方法的处理效果要优于传统的直方图修正算法，而其处理速度要快于各种自适应直方图修正方法。关键词:超分辨率超分辨率成像图像配准行处理方法微变焦二维位置测量直方图修正第Vlll页更多还原

【Abstract】 Image superresolution (SR) processing is the technology to reconstruct high resolution and high quality image(s) from a group of warped, blurred and noised low resolution images about the same scene. It breaks through the resolution limit of image acquisition equipment and can achieve data fusion on pixel level. Image SR processing has proved to be useful in many practical applications, such as remote sensing, military detection, medical imaging, machine vision and public security, etc. It can greatly improve spatial resolution and enhance image quality. So it is beneficial for computer to process, analyze and understand the image.The main contents and contributions of this dissertation are as follows:1. Spectral and spatial domain mathematic models of the SR problem are analyzed and established. A hierarchy image registration algorithm based on optical flow is given. A conditional row action iterative method is put forward. It fuses orthogonal projection and parallel projection and speed up the de-aliasing processing.2. Two spatial domain SR processing algorithms are proposed: Real-time frame iterative back-projection algorithm (RFE3P) and modified Cimmino fast row action iterative algorithm. By constructing a virtual panorama image and corresponding weight matrix, iterative back-projection (IBP) algorithm can be implemented from frame to frame in RFIBP. So video sequence can be processed real time. According to the modified Cimmino fast row action iterative algorithm, bilinear interpolation image blurring model on pixel level, which is a set of large sparse unstructured linear equations, is established. In order to solve the equations, a new row action iterative method based on Cimmino algorithm is proposed. The denominator, which is used to calculate the average iterative modification value in Cimmino algorithm, is replaced from number of rows to number of nonzero elements in a column of the system matrix. Since only is the valid projection taken into account, the algorithm converges quickly and greatly spares the computation memory. Moreover, parallel processing could be adopted if necessary. By combining SR and image mosaics, both the methods can ont only improve resolution of image, but also expand visual field.3. Rotate birefringence crystal method for micro-scanning image capturing is proposed. Its principle and machinery are discussed in detail. The relationship between the micro-translation in super-resolution imaging and reconstructing of high resolution image is analyzed. And it is proved that equal interval employed by micro-scanning is optimal but not absolutely needed. Unequal interval micro-scanning can also achieve super-resolutionimaging by controlling the distribution of micro-translation and increasing properly observed image frames. Therefore, the precision requirement to micro-translation machine is decreased significantly.4. A new vision measurement method based on image registration, which is used to measure 2-dimensional translation and rotation parameters of moving platform, is proposed. Benchmark graphics and lens are fixed, and CCD translates and rotates along with platform in imaging plane of benchmark graphics. Then the position of CCD in the whole imaging plane is located by image registration. Experiment result shows that system measurement precision can reach sub-micron level.5. The principle of micro-zooming is analyzed in spatial domain. Micro-zooming is the method that changes imaging focal length of optical system gradually to obtain a series of images about the same scene with different amplified multiples, and reconstruct high resolution image from these images. It can be used to realize superresolution imaging of opto-electronic imaging systems. Micro-zooming can be considered as the change of the area and position of sensor element’s corresponding to object radiance distribution region. So, a spatial model is established and the relationship between observed images and ideal high resolution image can be expressed by a set of linear equations. The high resolution image can be o更多还原