Abstract:

A new optical remote sensing image registration method signal gridding-scale invariant feature transform (SG-SIFT) based on signal theory and gridding is proposed. According to relationships among the image layers in the difference of Gaussians scale space, the feature points' number of each image is set in proportion to make the their distribution uniform in the scale space. In addition, a regular gridding method is introduced to achieve the well distribution of feature points in the image space. Then, error matching pairs are eliminated by a correspondence error checking. Statistical and visual results show that SG-SIFT is superior to standard scale invariant feature transform (SIFT) according to the feature points distribution, while the number of correct matching pairs from SG-SIFT is 17.47% more than that of uniform robust-scale invariant feature transform (UR-SIFT) in average and the evaluation indicator of root-mean square error confirms its superior performance to SIFT and UR-SIFT.

Key words: registration, optical remote sensing image, scale invariant feature transform, feature points distribution, feature matching