An Image Matching Algorithm Based on Singular Value Decomposition

Image matching plays an important role in many fields, such as computer vision, remote sensing and medical image analysis. Traditional-correlation-based matching methods require heavy computation time and they are sensitive to image rotation. In this paper, a new image matching algorithm based on singular value decomposition is proposed to efficiently match image pairs with arbitrary rotation angle. Corner points with dominant orientation are firstly extracted as feature points in the two images. The initial set of feature point matches is then obtained by singular value decomposition of a correspondence strength matrix. A new expression of the correspondence strength matrix is introduced to handle more complicated imaging conditions. Each element of this matrix is computed from the similarity measure between two feature points based on normalized cross-correlation. The similarity measure is invariant to image rotation by taking into account the dominant orientation of the two feature points. Finally the epipolar geometry constraint is imposed on the initial matches in order to reject the false matches. Experimental results on complicated real images show the effectiveness and robustness of the algorithm. Moreover, it is easy to combine the proposed algorithm with other matching techniques. Experimental results with SVD-SIFT prove the expansibility and practicability of the algorithm.