Research On Image Registration Algorithm

Registration is a fundamental task in image processing used to matchtwo or more pictures. Over the years, a broad range of techniques have beendeveloped for the various types of data and practical problems. In thepractical application, each kind of image registration method has its ownadvantages over the others. However, there still doesn’t exist a registrationalgorithm whose versatility is good enough to be applicable to all types ofimage registration. Therefore, seeking a kind of algorithm of high accuracyand robustness as well as good versatility to be applied in various fields hasbecome an important direction of our research.Image registration measurement can be classified into three categories,gray-based, Fourier-based and feature-based. In this paper, we choosefeature-based method as our major target. We first introduce the SIFTregistration algorithm, it is shown from a lot of tests and study that thealgorithm has a bad performance on the alignment of the two images in thepresence of perspective transformation. Then, we study the smoothlyvarying affine stitching (SVAS), the algorithm is able to handle the parallaxand solves the problems of SIFT registration algorithm to some extent.However, as for the SVAS algorithm, there are still some deficiencies suchas the long time-consuming and the unstable registration result. Weproposed three strategies for smoothly varying affine stitching to improveregistration performance based on the problem analysis, especially for theanti-parallax case. An effective way is first given to select proper feature toreduce the condition number of the matrix involves in the cost function toimprove the efficiency and stability of the SVAS, and then nonlocal meansfilter is employed to solve affine parameters. Finally, a specified gray interpolation method is designed to complete high quality image registration.With these ameliorations, the proposed algorithm significantly reduces thetime-consuming and produces better results than the ones from the SVASmethod. Experiment results verify that the proposed algorithm is robustagainst anti-parallax and yields satisfying image registration rapidly. In aword, the proposed method is of good applicability.The paper put forward a stereo matching method, we first introduce thefive basic constraints of stereo matching, and then we mainly focus on thecontinuity and compatibility to get geometric compatibility based on therelationship between adjacent points and the total disparity gradientconstraints. Finally, we combine the improved SVAS and the twoconstraints to be applied in stereo matching. As the improved SVASalgorithm matches features as a set, rather than individually, there is reduceddependency on feature descriptor uniqueness. This can modify the lowermatching probability of traditional stereo matching algorithm due to theinaccurate feature descriptors. The improved SVAS algorithm also solvesthe feature points’ transformation parameters to get more accurate matches.It can be used for feature matching for its relatively high matchingprobability and precision, the combination with the two constraints canfurther improve the matching performance to form an effective stereomatching algorithm. At the end of the paper, experiment results verify thesuperiority in the matching probability and precision of the proposedmethod.