The Study Of Image Stitching Algorithm Based On Feature Points

Digital image registration technology matches two or more images of certain scene from different viewpoints or sensors,at different times,or under diverse conditions.The quality of image registration would affect the following processing steps directly.There are three main methods in this field: transform-domain-based, feature-based and gray-based. Although they have their own advantages, none of them can deal with all the problems. In practice, a proper method is chosen at the balance of speed, accuracy and robustness. The dissertation describes feature-based methods deeply, and a novel image registration algorithm is proposed.The traditional corner-based algorithm was found to be sensitive to rotations, color difference and noise, and an automatic image mosaic method based on improved Harris feature points detector algorithm and Pseudo-Zernike moments were proposed. Major research works are as follows:1. Harris is simple, practical and running stability corner detection algorithm. When the corner point information is significant, the algorithm is high precision. But this algorithm uses Gaussian filtering method in processing the image. So the algorithm is relatively slow speed of operation, the information extracted feature points exists the phenomenon of the position offset, information lost and clustering. This paper proposes targeted B-spline-based variational filtering and fast search method for local window combination Harris corner detection algorithm.2. Image registration is a key step in the image mosaic, the general idea of image registration algorithm is that it is not only to ensure the accuracy of registration, but also to ensure that not too much computation and not be too complicated for implementation. this paper, the improved Harris corner detection algorithm and Pseudo-Zernike moments calculated image registration algorithm (Harris Feature Points Pseudo-Zernike Moments Algorithm). Firstly, the improved the Harris operator extracted image features; and then calculate the characteristics points'Pseudo-Zernike moments of the center neighborhood window, each feature point by comparing the Pseudo-Zernike moments neighbor Euclidean distance to obtain the initial matched feature points; Last the spurious feature points pair were rejected by RANSAC algorithm and to build affine transformation model between the images, and calculate model parameters, to achieve image registration. 3. The weighted average image fusion method is not very satisfactory for processing image fusion seam, sometime, appears shadow. In order to eliminate image stitching trace, here, we use a Cap weighting function for image fusion, to obtain good mosaic image.The experimental results show that the improved Harris algorithm detected feature points were uniform, accurate and rapid,and the proposed method of image mosaic is robust to rotation , translation and noise.