Image Registration Techniques Based On Scale Invariant Feature Transform (sift) Study

Image registration is a process, which gets two or more images to match in different times and different view angles with different sensors or same sensor. It is also an essential premise of image analysis tasks, such as image fusion, image mosaic, superimage restontiong etc. When image registration used in practice, there will be lots of reasons that cause image distortion, now available image registration techniques is hard to fit image great changes like scale, rotation, view etc. In particular, its method are complicated and accuracy rating is not high. So image registration is a hard work, the research of its skills has profound theoretical and practical significance.This paper makes a brief introduction about theory and background of image registration, summarizes its methods, and analyzes focus and trend of it. Now, the advantages of this methods have high rate of accuracy, less noise disturbance from the External world and suitable for wide applications, which is the dominant flow direction currently. The author carried on a series of studies and experiments from image registration based on feature as the main starting point on SIFT algorithm. Result of the experiments proves algorithm is invariant to rotation, scale and illumination of the image, and is also robust in view angle change, affine changes or noises disturbance.Although SIFT algorithm has numerous advantages, but algorithm need to search multi-scale space, and do much histogram weighted operations on lots of feature points. The time of registration will increase geometrically when meet big resolution images。In addition, there are many unstable points existed in SIFT feature points, which will reduce the efficiency and precision of registration. Based on above problems, this paper suggests the improvement of it:Considering SIFT feature point is based on pixel neighboring of gradient variation. The pixel keep large gradient variation can get more distinctiveness and contain more image information. Just corner point keeps large change on pixel neighboring variation. So this paper introduces Harris corner detect operator into the process of image registration based on SIFT algorithm, perfects its method, picks out Harris operator from these feature points when checking them in multi-scale space, chooses them that stands for image information, and it also reduces calculated amount when describing SIFT features. Experiments result shows that lots of feature points keep no distinctiveness is deleted. Retained feature points distribute in outline of the reference object and about half feature points be reduced. While the remained feature points can be better represented image information. Because the reducing of feature points, it also reduces calculated amount in follow steps. Analysis of experimental data prove that the proposed algorithm is retains the inherent advantages of SIFT algorithm. Besides, the registration speed and registration accuracy rate is improved.The last part of paper looks ahead the research direction on the field of image registration.