Image registration by integrating similarity and epipolor constraints

Accurate image registration plays an important role in computer vision, three-dimensional information acquisition and remote sensing processing. Recent researches on image registration are mostly focused on those by using either epipolar geometry or feature extraction, restricting the performance of image registration. A novel algorithm integrating similarity constraint and epipolar constraint is proposed in this paper. This algorithm firstly realizes the similarity constraint utilizing the Harris-Laplace detector combined with the SIFT (Scale Invariant Feature Transform) descriptor. Then, new approach of two-way epipolar constraint is proposed and used in the following processing, which satisfies the epipolar constraint and uniqueness constraint simultaneously. Thirdly, the iteration process is conducted until all constraints are satisfied. Additionally, a generalized algorithm for the situation without localization information of the image acquisition is given. The algorithm is examined through extensive experiments on various images from video sequence to digital camera images and the results show that the proposed algorithm exceeds SIFT in the aspect of match accuracy at the cost of a little more time consume.