The Research Of Image Stitching Based On Invariant Feature Matching

Image Stitching is a technology that alignment some images which could be overlapped each other, and finally produces a big eyespot image. It has an important applied value in the fields of medical image analysis, synthesized of the panoramic image, video compression and so on. So, image stitching is a research focus in the field of image processing. Because of its high robustness, image stitching based on the invariant feature matching has been the main method.This thesis research algorithms that related to image stitching based on invariant feature matching, and the following are the main works:(1)Rearch on the algorithm for extratation of image invartant feature. Firstly, this thesis introduced two different invariant region detectors:SIFT and Hessian-Affine. Then, local descriptors SIFT and CS-LBP is presented. Because coding’s robustness of CS-LBP is not very strong, this thesis improves CS-LBP by the fuzzy mearsure strategy of FLBP and names the impoved algorithm after CS-FLBP. The experimental results with Oxford University’s images matching set show that CS-FLBP performs favorably compared to SIFT and CS-LBP. Because the measure by comparing the distance of the closest neighbor to that of the second closest neighbor is an effective feature matching method, the best ratio of distances of closest and next closest is suggested for CS-FLBP.(2)A new algotithm based on the improved MSSE and the degree of concentration is proposed for the estimation of homography matrix. The candidate matrix can be got by sampling data from matched pairs. MSSE assumes the distribution of residuals of the correct matched pairs’s fitting the correct matrix is a normal distribution. However, in the real data, the distribution is not smooth which lead to MSSE’s incorrect estimation of the inlier/outier separation threshold. To overcome MSSE’s shortcome, an improved strategy that regards the discontinuous point has a maximum of following discontinuous points as separate point of residuals is proposed. Standard deviation of inliers’ residuals got by the separate point is used to compute concentration. Then, this thesis designs an algorithm based on concentration for the estimation of homography matrix on the condition of inliner is known. The experimental of image registration show that the algorithm performs favorably compared to LMeds and LQS. Furemore, the algorithm is improved by the fact that the size of intersection of inliers estimated by the sample of correct homography matrix. The new algorithm can determine the inlier rate by itself in the execution, and the experimental of image registration show that it gets more accurate homography matrix compared to RANSAC algorithm with best threshold.(3) Design and implement the algorithm of images’automatic stitching based on invariant feature matching. This thesis used bundle adjustment algorithm to solve for the error of global homography matrix by multiple passes. For get panoramas with higher quality, the brightnesses of the images are firstly adjusted and then we blend images by multi-band algorithm.