A Novel Image Registration Algorithm Using SIFT Feature Descriptors

This paper focuses on the problem of image registration, which is a fundamental and important problem in computer vision. As local features are effective to tackle the task of image registration, hence, we exploit SIFT feature descriptors to describe the interesting key points in images. The reasons is that SIFT can extract features which are invariant to scaling, orientation, affine transforms and illumination changes. Afterwards, to lower the effects of outliers, the nearest distance ratio method is used to get initial matched feature descriptor pairs, and then scale-orientation joint restriction is used to detect false matching SIFT descriptor pairs. After using Random Sample Consensus to prune outliers, image registration results can be obtained. In the end, we develop an experiment based on QuickBird and IKONOS datasets, and very positive results are achieved.