Abstract: The method for remote sensing image registration based on scale-invariant feature transform(SIFT) has the advantage of hig haccuracy and good stability.However, the method is very time-consuming because of the large size of the image and the huge quantity of feature points.This paper presented a parallel acceleration method for very high resolution remote sensing image registration which builds the Gaussian pyramid by hardware implementation on GPU.We used the shared memory to cache the temporary extremum at high speed when identifying the keypoint,which effectivelydecreases the time for the keypoint extraction.Meanwhile,we divided the whole image into blocks and used OpenMP to match the feature-points and build parallel acceleration of the affine model.Compared with the traditional registration method——SIFT,this method is 3 times faster.We concluded that the runtime of the keypoint extraction has linear relationship with the quantity of the keypoints,and the acceleration ratio raise with the density of the keypoints going up.

Key words: GPU,Remote sensing image,SIFT,Registration