Harris feature and coherent point drift based remote sensing image registration

Abstract

Non-rigid point-set registration based image registration is a technology frequently used in image retrieval, stereo matching and the analysis of satellite and medical images. In remote sense image analysis, Harris corner detection is often chosen as an ideal approach of feature extraction. The method we propose utilizes the feature metric produced by Harris corner detection, which is not employed in current methods, and integrate it into the Coherent Point Drift framework to enhance accuracy. We first construct a likelihood descriptor of point-to-point correspondence, then this likelihood value is used as a prior probability term in the Gaussian mixture model. Finally, we use the Expectation Maximization algorithm to iteratively match the points. Our contribution includes finding a way of normalizing the feature metric data and constructing a proper descriptor to incorporate the Harris feature metric and the Euclidean distance which can minimize the negative effects of the deviations in the feature metric values. Experiments are conducted upon remote sense images, compared against four state-of-the-art image registration algorithms, including two non-iterative methods and two iterative methods, where our method show the smallest error rate and the best registered image result.