Abstract
An accurate image registration is a fundamental stage in many image processing problems. In this paper, a new and fast registration approach based on scale invariant feature transform (SIFT) key-points, under Euclidean transformation model, is proposed. The core idea of the proposed method is estimation of rotation angle and vertical and horizontal shifts using averaging of differences of SIFT key-point pairs locations. The method is simple but requires some tuning modules for accurate estimation. Orientation modification and compensation and shift compensation are some of the proposed modules. The proposed method is fast, about ive times faster than RANSAC method for model parameters estimation. The accuracy of the proposed method is compared with some popular registration methods. Various comparisons have been done with LIVE database images with known motion vectors. The experimental results over two real video sequences show the high performance of the proposed algorithm in a super-resolution application.
The authors are indebted to anonymous referees for valuable comments. We would also like to thank Dr Patrick Vandewalle (Laboratoire de communications audiovisuelles, École Polytechnique Fédérale de Lausanne, Switzerland), for providing us with his super-resolution package, Dr Alan Brooks (Department of EECS, Northwestern University, USA, http://dailyburrito.com/m-files/) for his stabilisation Matlab codes and Professor David Lowe for his SIFT program. They also thank to Dr Peter Kovesi (http://www.csse.uwa.edu.au/_pk/research/matlabfns/) and Dr Hamid Sheikh for providing LIVE datasetCitation18, ”Dr Gh. Mohajeri Minaei for video sequence 2„ and A. Ghodsi (PhD Student in Spatial Processes, University Putra Malaysia) for his contribution.