A fast and robust image stitching algorithm

This paper proposed a coarse-to-fine method to construct mosaics from images. A two-stage feature-based robust estimation method is used, which quickly searches for a small number of consistent matches defining a transformation that is well approximation to the correct one. After getting all matching pairs, geometric relations of neighboring points are used to eliminate the false matches, and then Levenberg-Marquardt's nonlinear method is used to minimize the difference and update the related transformation. At last, panorama blending with two-band blending method gains high quality in spite of large lighting variation between images. Results showed that the proposed method built high-quality panoramic image mosaics in high speed.