The Research On Multi-View Image Registration

Digital image registration is a fundamental problem in computer vision. It has a wide range of applications in three-dimensional image reconstruction, object recognition and classification, self-tuning of the camera and so on. Digital image registration is usually a pre-processing stage in digital image processing, such as digital image fusion and mosaicing. The technology of digital image registration can match different images which are obtained from a same scene in different conditions. It produces a new interpretation of this scene, and this explanation can not be obtained from a single image.In order to obtain more detailed and rich information of image, we often shoot the same scene from different angles. By this shooting we get a number of different images which are different in views, that is, multi-view image. This work is to research the problem of matching images which are got from the same scene in different views, that is,the problem of multi-view image registration. The research of multi-view image registration has an important meaning and value to promote the development of technology of digital image registration.In this work, a new method is proposed to match the multi-view image. Firstly, I use the method of affine invariant detection which is derived from the method of Harris feature detection to extract the feature points of multi-view image. Secondly, I normalize the invariant feature regions which are the neighborhood of the feature points in order to change the multi-view image registration issues into image registration issues of a simple rigid transformation. Thirdly, I use an improvement of the SIFT descriptor to describe the feature points. Finally, I use the function of the distance to calculate the similarity in order to match feature points.In my work I experiment this method and the SIFT method separately to match the multi-view image. Experiments show that this approach can get a better results of matching than the SIFT approach when the image has a larger transformation of views.