Feature matching is a difficult and key problem of the feature-based image registration methods. A new algorithm based on feature similarity is proposed in this paper to deal with the correspondence of features. It mainly consists of three steps: first, a candidate matched feature pairs set is obtained using feature similarity which is computed by the invariant descriptors of features extracted from the reference and sensed images; second, the minimum number optimal matched feature pairs, which can be used to estimate the parameters of geometric transformation model between the two images, are acquired from the candidate set via an exhaustive search strategy; third, all the matched features are estimated according to the nearest neighbour rule. The characteristic of the algorithm is that it combines the methods using the criterion of the minimum distance classification, row and column matching likelihood coefficients algorithms and can overcome their shortcomings. Experiment results are very encouraging and prove that the proposed method is comparatively robust and effective.
展开▼
