基于特征库的遥感图像自动配准研究

【作者】 袁海军；

【导师】 李小文； 刘强； 顾行发；

【作者基本信息】 电子科技大学 ， 模式识别与智能系统， 2007， 硕士

【摘要】 中巴资源卫星02星(CBERS02)是我国与巴西联合研制的陆地资源卫星,搭载了CCD、IRMSS和WFI三种传感器,自2003年10月21日发射升空以来,已经传回了大量的遥感图像。然而,由于图像的细节模糊,图像数据量大,图像之间的配准处理不方便,使得CBERS02的CCD图像的应用受到一定的制约。因此,本文以CBERS02的CCD图像为研究对象,围绕图像配准,开展了如下几个方面的研究:1.通过对典型配准方法的总结和比较分析,认为基于特征的配准方法具有较好的灵活性,同时在配准过程中只需要处理从图像中提取的特征,而使得运算的数据量减少,算法的效率比较高。因此,在本文的研究中就采用了这种方法。2.基于特征的配准方法的核心就在于从图像中提取特征和进行特征匹配。在图像所具有的多种特征中,本文以兴趣点特征的提取为研究重点。为了选取对CBERS02的CCD图像具有最好适应性的算法,本文提出了基于典型兴趣点的算法准确度和相对重复度两个评价标准。结合算法的运行速度,通过定量的比较,基于互相关的PCD算法成为本文的选择。由于PCD算法在兴趣点定位能力上的不足和CBERS02的CCD图像本身具有细节模糊的特点,本文提出了兴趣点丰富区域特征(AAIP),以改善特征的定位精度。实验结果表明了从图像中提取AAIP特征的合理性和有效性。3. AAIP特征本质上是区域特征,因此区域特征的匹配成为本文的一个研究重点。本文首先分析了几种典型的区域特征匹配算法,并讨论了可能的改进。为了选取对CBERS02的CCD图像具有最好适应性的算法,本文从相似度的尖锐度、运行速度、抗旋转性和对图像灰度的依赖性四个方面进行了比对实验。对实验结果进行定量分析和比较,基于互信息的匹配算法成为本文的选择。4.为了提高图像配准的效率,重复利用图像配准过程产生的特征,本文利用数据库将图像配准过程中产生的特征以有效的形式保存起来,形成特征库,当需要进行新的配准处理时,首先根据目标图像的经纬度信息,从特征库中选取特征,形成参考图像,然后将目标图像跟参考图像进行配准。由于特征是已有的,所以只需要根据本文的匹配算法从目标图像中寻找到相应的特征,建立同名点,进而完成配准。这使得配准处理中,特征提取这一最耗时间的过程被跳过,所以配准的效率变高了。同时,这也扩展了配准的概念,将图像配准从两幅图像之间拓展到了图像与特征库之间。实验结果表明了这一算法的有效性,配准的精度为1个像素。总之,本文的研究表明:基于PCD算法的AAIP特征提取和基于互信息的特征匹配对于CBERS02的CCD图像具有较好的适应性,基于特征库的配准方法对于CBERS02的CCD图像是具有可行性的。更多还原

【Abstract】 The second China-Brazil Earth Resource Satellite (CBERS-02) was developed by China and Brazil, onboarding three kind of remote sensors: CCD, IRMSS and WFI. Since it was successfully launched in October 2003, it has produced a large amount of remote sensing image. But the image blurring and the large of image makes the registration which is necessary in the process of remote sensing image application difficult. In order to find or design a suitable scheme to complete the image registration effectively and efficiently, the following job have been done based on the CBERS02’s CCD image.1. Based on the summary and compare of classical image registration algorithm, we conclude that the registration algorithm based on feature (RAF) has two aspects of advantage: (1) it is flexible; (2) it is high efficient because the computing based on feature only use a little part of the data. And so in this paper we will take advantage of this algorithm.2. The key part of RAF includes feature extraction and feature match. The image includes different kinds of feature, in this paper we focus on interesting point (IP) extraction. In order to find the most suitable IP extraction algorithm for the CBERS02’s CCD image, two new criterions—the algorithm accuracy and relative repeatability based on classical IP is developed. Combined with the computing speed, the Plessey Corner Detector (PCD) based on correlation is a good choice based on quantitative compare. Because of the fault of PCD on localization and the blurring of CBERS02’s CCD image, a new kind of feature—Area with Abundant Interest Points (AAIP) is developed in order to improve the localization. The reasonability and validity is proved by the experiments.3. In fact, the AAIP is area feature, and the area feature match becomes another research focus. Firstly some kinds of classical match algorithm based on area feature is discussed and compared, and some impossible improvement is developed. In order to find the most suitable algorithm, experiments are carried out to learn about the similarity’s acuity, computing speed, anti-rotation and dependence on gray of different algorithm. Based on quantitative analysis of the result of the experiment we conclude that the algorithm based on mutual information is the best choice for our image.4. In order to advance the efficiency of the registration and reuse the feature extracted in the process of registration, we store the feature to form a feature database. When we carry out a new registration, firstly we select some feature from the database according to the latitude and longitude to reconstruct a reference image, and then start the registration. Because the feature exists before the registration, feature extraction which is the most time-consuming doesn’t need to be carried out, and the efficiency of the algorithm is advanced. At the same time we extend the concept of the image registration from image-to-image to image-to-feature database. The experiment result proves that this way is effective, and the accuracy of the algorithm is about 1 pixel.In a word my dissertation demonstrates that the extraction of AAIP based on PCD algorithm and the feature match based on mutual information is suitable for CBERS02’s CCD image; that the registration scheme based on feature database is feasible.更多还原