Image matching based on orientation–magnitude histograms and global consistency

A novel image matching method based on the gradient space is proposed. Image pyramid combined with the Hessian matrix is used to detect scale-invariant interesting points. A new descriptor i.e. an orientation–magnitude histogram is introduced to describe the image content around an interesting point. The proposed local descriptor is proved to be invariant to image rotation. Since the matching result based on the similarities of the descriptors of interesting points always contains outliers, a steepest descent method that optimizes the global consistency of interesting points is presented to remove false matches. The experiments show that the proposed approach is invariant to rotation and scale, robust to the variation of focal lengths, illumination change, occlusion, noises and image blur. Our approach shows better performance than SIFT on multi-view and affine-transformation images. The application of the proposed method to image registration exhibits a good result.

► A novel image matching based on orientation–magnitude histograms is proposed.
► It is invariant to scale and rotation, robust to various imaging conditions.
► Our method outperforms SIFT on multi-view and affine-transformation images.
► Its application to image registration shows a good result.