GPU-accelerated large-size VHR images registration via coarse-to-fine matching

This paper presents a coarse-to-fine large-size very high resolution image registration method. This method uses compute unified device architecture to speed up the acquisition of control points and image rectification. In coarse registration, the scale-invariant feature transform algorithm is used to match control points to estimate the initial global transformation parameters between images to be registered. The initial parameters are then used to guide Oriented FAST and Rotated BRIEF (ORB) feature matching in fine registration. To fix local distortion, image rectification is based on a linear mapping function computed from triangulations, and a self-adaptive scan filling algorithm is proposed to determine which triangle each pixel belongs to. Experiments are conducted with large-size satellite images. Results show that a graphics processing unit can obtain significant-acceleration factors while maintaining registration accuracy.