Researching Of Image Stitching Algorithm Based On Feature Points

Image stitching technology is getting a series of the small image which have some overlap region into a panorama image.This technology is widely used in 3D reconstruction,interactive panoramic image preview,multi-node movies and some other technologies which use the 3D model to construct reality environment image.According to the different principles of registration,image stitching technology can be divided into many different methods.But generally there are three steps including: image preprocessing,image registration and image fusion.Image registration is the foundation and core step of image stitching technology,so the development of image stitching technology mostly depends on the innovation and optimization of image registration technology.The image registration based on SIFT(Scale-invariant feature transform)is a classic method of image registration based on feature points.SIFT algorithm has better robustness,the feature points it found is not going to disappear while with the changing of illumination and the noise in the images.Its robustness is well for the illumination and the noise in the images,these points always be a corner points and edge points and dark areas in the bright and bright regions in the dark,and the feature points thus obtained is stable,and can reduce the amount ofcalculation in the process of registration,it can be convenient to complete the image registration.Based on the existing research results at home and abroad on the basis of study and research,this paper mainly completed the following several aspects:(1)We analyze the image matching technology in theory,and introduce the basic knowledge and process of image matching technology.We introduce in image matching technology including image preprocessing,image transformation,image registration,image fusion etc.(2)We introduce the state of the art of image registration method based on SIFT feature points,and then we improve and optimize a registration method,which is called smoothly varying affine stitching.We find the affine transformation matrix which relates to each of the original image feature points,can adapt to a variety of different types of scenarios better,it can find more precise matching points in constant iterative operation.Based on the deficiency of this approach,this paper gets some optimizing,with the method of the projection matrix to describe the transformation between pixels instead of the affine matrix.Experiments show that the projection matrix of smoothly varying stitching is better than the effect of the affine matrix,and can adapt to more types of images.(3)In this paper,we compares the effect of the blending of several different ways,taking the best optimal blending method.At the end of the paper,on the base of Poisson blending,we propose and use a Poisson blending method based on cylindrical projection.Finally,with the method of simulation experiment compares,experiments show that this kind of blending way to get more smoothly stitching images,and the detail processing is better.