【作者】 陈锦伟；

【导师】 徐之海； 冯华君；

【作者基本信息】 浙江大学 ， 光学工程， 2014， 博士

【摘要】 遥感图像是指从远距离平台上利用光电成像载荷获取的地物目标图像,常见的平台如飞机和卫星。由于其覆盖范围广,光谱频段丰富,在军事和民用上都起到了巨大的作用。但是随着人类活动的区域日益广泛,对于遥感图像的整体要求也越来越高,希望在获得高分辨率图像的同时,其覆盖的区域更加的宽广,以至于能对一个较大的区域进行细致深入的分析研究,为后续的人类决策提供支持。但是,受限于目前的技术水平,高分辨率和宽视场仍然是一对矛盾体。常见的解决方法是通过图像配准和拼接的方法,也就是是利用传感器得到多幅高分辨率的小视场图像,然后对这些具有一定重叠区域的图像进行配准拼接,得到一幅大视场图像。这种矛盾在敏捷卫星的任务中会出现的更加频繁,敏捷卫星是近年来出现的新型卫星,具有优秀的姿态加速功能,使得敏捷卫星可以对目标进行更快速的瞄准和更精确的扫描,大大提高了满足复杂任务的要求。但是敏捷卫星由于其高机动性,会使得成像环境更加的复杂,从而影响最终的遥感图像质量,因此对敏捷卫星的成像模式进行深入细致地研究,可以定量化地分析机动成像后的图像质量变化特性,而因为敏捷特性带来的图像配准拼接问题也会成为更大的挑战。通过对经典的配准和拼接算法的研究,为后续的敏捷卫星的图像拼接约束分析提供理论支持。图像配准拼接技术可分为两类：基于区域的图像配准技术,如序列相似性检测算法、交叉相关相似性度量函数和傅里叶变换算法；基于特征点配准算法包括sIFT特征点提取算法和Harris角点提取算法。对这些配准算法的研究可以从图像角度对敏捷模式带来的影响进行分析,同时也为基于遥感景物内容特性的配准算法提供了基础。通过构建几何模型、辐射模型、相机模型和大气模型,来模拟分析敏捷卫星的机动成像模式,并且提出梯度信息熵、仿射退化度和立体结构相似度等新颖评价指标来预估图像质量。基于四个模型对敏捷卫星的成像模式进行分解,使得可以通过卫星轨道等基本参数来定量地分析卫星图像的各个参数,比如分辨率、幅宽,以及后续的图像质量指标。在理论模型的基础上,编写了“敏捷卫星成像仿真和质量分析“软件,可以对敏捷卫星进行各个模型的分析,同时可以对图像退化进行仿真模拟。利用图像复原和图像融合的预处理技术,改善敏捷成像下可能带来的图像模糊和颜色退化现象。针对图像复原技术,首先分析了典型的复原算法,然后提出了基于FOE模型的图像复原算法,最后研究对比了图像复原技术对于特征点提取技术的影响。针对图像颜色退化,本文分析了典型的图像融合算法,并对融合后的遥感图像进行配准拼接,提高其目标识别的能力。通过对遥感图像景物内容特性的分析,提出了几种改善性的配准和拼接算法,使得遥感图像有更广泛的实际应用能力。首先提出了一种基于梯度信息权重优化的配准技术,利用图像的梯度信息对图像的特征点进行权重划分,利用权重值对特征点进行区别优化,最后将联合的图像拼接技术应用到重叠区域,实现对敏感区域进行高精度匹配的目的。针对遥感图像含有丰富的内容特性,设计了一种双特征点配准算法,利用SIFT和Harris短发提取特征点种类的区别,分别对角点区域密集的地方实现Harris特征点提取,对于相对平坦的区域(比如草地和水面)进行SIFT特征点提取,最后得到可以应付复杂环境下得到的推扫图像。遥感图像通常含有丰富的内容,使得其特征点数量通常非常的巨大,应对这种情况,提出了基于精炼控制点的配准技术,提取少量的精确特征点可以减少错误匹配特征点对结果的干扰,同时可以降低对遥感图像的处理难度。更多还原

【Abstract】 Remote sensing images which are always acquired from these devices on plane or satellite have been widely used in People’s daily life. We always hope to get high resolution and large size remote sensing images which can help to analysis of a region more deeply. However, it is contradiction between field of view and resolution in remote sensing field. The common solution is to acquire final image step by step. Firstly, we acquire small images which have high resolution. Secondly, we utilize registration and blending techniques to stitch two or more images which have overlapped areas into one large image. The registration and blending are very important techniques which can improve practical performance of the image. New agile satellites are appeared in recent years which have excellent acceleration ability. Agile satellite can aim and scan targets more accurately which make it meet the needs of complex tasks. However, due to its high mobility, agile satellites will make the imaging environment more complex, which affect the final quality of remote sensing images. We study the agile model of satellite deeply and in detail, quantified analysis the change of image quality after agile model.To do a research summary of a variety of registration techniques of images, such as Sequential Similarity Detection Algorithm (SSDA), Cross Correlation (CC) and Fourier transform. The Scale space can help to Understanding of scale-invariant principle. Scale Invariant Feature Transform (SIFT) and Harris corner extraction are very famous methods which can lay a solid foundation for registration and stitching of remote sensing images.Agile satellites have good mobile abilities which can get more different parts of the image. We firstly build geometric model, radiation model, atmospheric model and camera model, secondly propose some principals:Signal to Noise Ratio (SNR), MTF, Gradient Information Entropy, Affine degeneration and Structure Similarity Parameter to evaluate the quality of the image acquired from agile satellites. Our models can evaluate the quality before the target which can help users to make a decision of the target.Image restoration and Image fusion can help to reduce image fuzzy and recover the color information of image. Field of Experts restoration method can improve the quality of blurring image, which can make blurring image be used in some situation where is very strict about the quality of the image. The restoration image can extract more feature points than blurring image, and the better matching precision will make the registration more accurate. Image fusion can combine panchromatic image and multispectral image which make the fusion image has high resolution and rich color information.We have proposed three improved registration methods of remote sensing images, which can help to acquire better adaption ability in practical application. Firstly, registration method based on human vision introduced the technology of feature point weights. Gradient information of image is the judgment standard which can realize registration of image automatically. The combined stitching method can make the transition of overlap of image more naturally. The double-feature registration method employs two feature extractions, SIFT and Harris. Harris can extract corner features more accurate in urban images, however bad ability in the regions such as water and grassland. After comprehensive comparison, we utilize the SIFT to extract feature from flat area, and Harris to extract feature from urban area. This combined method can acquire more accurate result. Registration based on refined control points can reduce affection of the wrong match of feature points. Remote sensing images always have a lot of objects, so we can extract a mass of feature points which however will also increase the rate of wrong matching. In fact, we only need a few points to resolve the transformation between two or more images. This method can help us to find five groups of well-matched feature points, which make the registration easier.更多还原