Study On Key Technology In Feature-Based Image Registration

Image registration is the process of aligning and fusion of images referring to the same scene taken from different times, different sensors or different perspectives. It is a hot spot on the computer vision, pattern recognition, medical image processing and remote sensing processing. Image registration is widely used in the multi-source remote sensing data integration and analysis, motion tracking of small target under complex scene detection, matching of landscape and map, image stitching and topographic height reconstruction. According to different classes of image registration, the methods can be divided into the area-based and the point-based. The image registration algorithms based on the image feature are widely used since they are robust and efficient when the illumination changes, rigid transformation or affine transformation occur.In this dissertation, we study some key technologies of feature-based image registration. We focus on these issues and take in-depth analyses and researches: feature extraction, feature description and transformation matrix estimation. The main achievements of this dissertation are as following:An operator based on Mexican hat function is proposed and an image registration algorithm using feature pionts grouping and matching is gotten. The local areas in the image are detected using the Mexican hat operator and matched preliminarily. Then the feature points are detected in the scale space using the operator. The feature points are grouped according to the mathed local areas, the points in a local area are matched with the points in the matched area. The mismatchings are rejected and the the image transaction function is gotten by the random sample consensus eventually.A feature point description method is proposed based on the local binary pattern. We classify the feature points using the univalue segment assimilating nucleus(USAN). The feature point satisfied the USAN rule is described using the multilayer uniform local binary pattern. The description method is called the univalue segment local binary pattern descriptor. The descriptor is simple and it reduces the dimensionality of the local binary pattern.The feature point description method based on the local structure of nearest neighbors can not descibe the featur point accurately. This method usually becaomes complex when it rejects the mismatching. To solve these problems, a novel feature point descriptor is proposed. We get a coarse matching of feature points using the distance voting,the mean of the inner point set is named the global reference point. We define the descriptor using the rank order encoding in the artificial neural network, the connective weight among feature points in the rank order enconding is defined based on the global reference point. The descriptor is named the global reference angle order operator. At last, an image registration metod based on the DAISY operator and the global reference angle order operator is gotten.The transformation matrix estimation of the image registration can be seen as a problem of objective function optimization. The objective function contains two parts of optimal solutions: the biggest number of matched point pairs and the image transformation matrix with the highest accuracy. We can use two kinds of feature point descriptors to achieve the two parts of solutions. The descriptor based on the grey level histogram can be taken to get the image transformation matrix, the descriptor based o n the space structure among points can be used for getting the matched point pairs. The improved objective function and its optimal solutions can make the traditional methods of image registration faster and more precise.