【作者】 王运；

【导师】 金光； 颜昌翔；

【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， 2011， 博士

【摘要】 超光谱成像仪遥感能同时获得目标的空间和光谱信息。本文研究的超光谱成像仪是航天对地成像系统,由于各种因素的影响,会造成超光谱图像存在着一定的几何畸变,超光谱图像不能真实反映地物的光谱信息;本文研究的超光谱成像仪采用两个光谱仪来拓宽对地成像的光谱范围,这种方式会造成光谱仪波段图像之间的失配变形,影响地物目标的光谱纯度,降低超光谱图像的质量;另外,高光谱分辨率的超光谱图像需与高空间分辨率的全色波段图像融合,以增强图像的空间分辨率,而高精度的图像配准是融合的基本前提。本论文主要针对超光谱图像的系统几何校正、光谱仪图像配准、超光谱图像与全色波段图像的配准展开研究,完成了以下工作:1.建立了超光谱遥感图像系统几何校正模型,利用仿真实验验证了该模型的可行性,并在此基础上构建了光谱仪图像失配模型,分析了垂直对地成像和沿轨摆扫运动补偿对地成像时光谱仪图像的失配变形,采用蒙特卡洛法分析了光谱仪图像在不同狭缝间距,不同星下点纬度时姿态角误差、指向镜角度误差对失配变形的影响,分析结果可以为超光谱成像仪狭缝设计以及光谱仪图像配准提供参考。2.根据超光谱图像数据的特点,提出一种结合波段选择和主成分变换的方法,该方法能将三维超光谱图像转换为灰度图像以便图像的配准。波段选择方法能保证从光谱仪图像中选择相似性高的波段组合,主成分变换方法能提高生成图像的信息量和信噪比。3.提出一种像素到像素的光谱仪图像配准方法,该方法首先通过局部图像窗口平移估计来计算图像的平移矢量场,然后利用估计的平移参数通过灰度插值来实现光谱仪图像的配准。其中局部图像窗口平移估计是该方法的关键,其估计精度决定了图像配准的质量,在相位相关方法的基础上提出了两种局部图像窗口平移估计方法:(1)在相位相关空域估计方法中,提出一种改进的sinc函数拟合相位相关面估计平移参数的方法,实验结果显示该方法在光谱图像的配准精度能达到0.03个像素,比sinc拟合函数方法提高2～3倍,说明了提出的拟合函数能更准确地描述相位相关面的实际分布。将该方法应用到真实光谱仪图像配准实验中,结果表明该方法有效可行。(2)在相位相关频域估计方法的基础上,提出一种基于核密度估计的平移参数估计方法,该方法将奇异值分解和核密度估计有机结合,奇异值分解能提高相位相关矩阵的质量,核密度估计能在受噪声干扰的相位分量中正确估计到平移参数。从对比实验来看,该方法比最小二乘法对噪声有更强的抗干扰能力,能有效提高平移参数估计精度和稳健性,在光谱图像的配准实验中精度能达到0.04个像素,比最小二乘拟合方法提高1倍。该方法可以做为上述空域估计方法的补充。4.提出了一种基于特征点的全色波段图像和超光谱图像配准方法。该方法分为两步,首先利用传统的基于特征点的方法对待配准图像进行粗配准,然后利用提出的特征点位置优化方法对特征点位置进行优化处理,利用优化后的特征点对应关系对图像进行二次配准,实验结果表明该方法能显著提高图像配准的精度,满足全色波段图像和超光谱图像0.3个像素配准精度的需求。更多还原

【Abstract】 Hyperspectral imagery can provide image data containing both spatial and spectral information. In this thesis, the Hyperspectral imagery was an aerospace earth observing system, because of the influences of various factors, the hyperspectral images would have geometric distortion that can’t real reflect the spectral characteristics of the objects on the Earth. Besides, the hyperspectral imagery collected data using two spectrometers to span a broad spectral range, this configuration caused an inherent misregistration between the two sets of spectral bands, the misregistration between spectral bands had detrimental effects on the spectral purity of each pixel which seriously degrade the quality of the hyperspectral images. Furthermore, to enhance the spatial resolution, the hyperspectral images were usually fusion with the high spatial resolution panchromatic image, the essential prerequisite for this was the precise image registration. Therefore, this thesis was focus on the method of the systematic geometric correction, spectral bands misregistration correction and the registration between the hyperspectral images and the panchromatic image. The accomplished research works as follows:1. A hyperspectral imaging systematic geometric correction model was established, the feasibility of the model was verified by the simulation experiment. Based on this, a misregisation model was established and the misregistration deformation of the spectrometer images was analyzed when imaging vertically and imaging with motion compensation. With the Monte Carlo method, the influences on the misregistration deformation under different slit interval and different latitude of nadir point were analyzed. These results can be a reference for slit design and image registration.2. Due to the characteristic of hyperspectral images, a method that combining the band selection and principal component transformation was proposed, which can convert 3-D images to grayscale image for image registration. The band selection method can guarantee the corresponding generated images to have the largest spatial similarity, and the principal component transformation can improve the information and SNR of the generated images.3. A pixel-to-pixel image registration method for spectrometer images was proposed. Firstly, the motion vector was calculated by local image motion estimation method. Then, the motion vector was used to register the spectrometer images by intensity interpolation. The local image motion estimation method was the key technique which determines the accuracy of the image registration. Two methods for local image motion estimation were proposed based on the phase correlation method:(1) A modified sinc fitting function to obtain a subpixel estimate of motion was proposed. The experiment results show that this new method can enable the estimation motion with 0.03 pixel accuracy , about 2-3 times higher than the sinc function fitting method. In the real spectrometer image registration experiments, the real results show that the method is feasible and efficient.(2) A subpixel motion estimation method combining singular value decomposition and kernel density estimation in frequency domain was proposed. Singular value decomposition can improve the quality of the phase correlation matrix, and the kernel density estimation can estimate the motion correctly from the phase correlation corrupted by the noise. Comparative results show that this new method is more accurate and robust than the least square fitting method in the presence of noise. The registration accuracy can reach 0.04 pixel in spectral images registration, about 1 times higher than the least square fitting method. This method can be used as supplementary motion estimation method of the above method.4. An image registration method Based on feature points between the panchromatic image and the hyperspectral images was proposed. The method consists of two steps: firstly, the images were roughly registered based on traditional method using feature points; secondly, a location optimization algorithm was used to improve the location of the Corresponding feature points, and then the image transformation was re-estimated. Experimental results show that this method can significantly improve the registration accuracy to meet the accuracy needs of 0.3 pixel between the panchromatic image and hyperspectral images registration.更多还原