Research For Image Registration And Its Application

The image registration technology is the basic task of the computer vision, whose application area include medical image processing, remote sensing image processing, image fusion, image searching, robot vision, virtual reality, object recognition, GIS and so on. The essence of the image registration is that computing the override of the two images to estimate the global space transformation relation. Depending on the characteristics of the application scene, the different application strategy should be taken. The image registration technology can be classified into two classes---pixel-based registration and feature-based registration, each with its own characteristics. The pixel-based registration also can be called direct registration. To use the direct registration, a suitable error metric must first be chosen to compare the images. But choosing a suitable error metric that can be used for all images is an impossible task, and also considering the high computing cost, so the direct registration just applied to several special areas, such as medical images processing. Along with SIFT and SURF had been proposed, the feature-based registration technology had overcome the shortcomings of the direct registration technologies. The feature-based registrations had become the most popular registration technology for its excellent performance in the last decade.The process of the feature-based registration can be divided into the three sub processes:key-points detection, descriptor construct and descriptors matching. The key points contain the location information that needed to construct the descriptor. And the descriptors that also called descriptor vectors used for discriminating the key-points. So it’s obvious that extracting stable key points and constructing the discriminate descriptors is the focal point of the study of the image registration technology.In this paper, the key points detecting algorithm based on nonlinear space and the local binary descriptor construct algorithms are the main contains. The key points are these points that have some or sort ability to discriminate themselves to the common pixels. The key points contain the key location of the descriptor, which to some extent have the ability to influence the matching result. In Gaussian scale space, the advantages of selecting coarser scales are the reduction of noise and the emphasis of more prominent structure especially at the high scale level. That reduction in localization accuracy is the price to pay. It seems more appropriate to make blurring locally adaptive to the image data so that noise will be blurred, but details or edges will remain unaffected. To achieve this, different nonlinear scale space approaches have been proposed to improve on the Gaussian scale space approach.In order to get the scale, rotation, and viewpoint invariant descriptors, the first step is to estimate the main orientation of the key points’ domain. In this step, the pixel pairs selected by the designed selecting model should be divided into two categories, one of them called short-distances pixel-pairs and the left one called long-distance pixel pairs. The ratio of them can be seen as the tan of the main orientation of the descriptor. And then the selected pairs normalized by the main orientation. At last, some pairs which cannot meet the demands of the descriptors are filtered by using the sampling model and threshold, and the left512pairs construct the final binary descriptors string.At the end of this paper, the image registration technology has been applied to the image mosaic and the object detecting. The purpose of that is to give some examples how to use the transformation matrix that got by the image registration technology.