Abstract: Aiming at the problems of dislocation and ghosting in the process of large parallax image stitching, an image stitching algorithm based on feature clustering is proposed. Firstly, the Tyson polygon is constructed in the overlapping area of the target image based on the distribution of matched feature points. Secondly, the improved AGNES hierarchical clustering algorithm is used to cluster the feature points, and merge the Tyson polygons represented by the feature points in the corresponding group, so as to get each sub-plane of the overlapping region of the target image. Finally, the homography matrix of the corresponding subplane is solved, and the homography matrix of the non-overlapping region is assigned according to the nearest principle, then the target image is transformed by projection to get the mosaic image. The experimental results show that the proposed algorithm has high registration accuracy and can effectively improve the misregistration and local distortion in large parallax image stitching.

Key words: image stitching, parallax, Thiessen polygon, clustering, homography matrix