Image registration is a central task in many image processing applications. We focus on multispectral or hyperspectral remote sensing images that are used, for example, to analyze changes in land use, or to monitor the environmental effect of natural disasters. In all these examples, different images of the same area are required. Feature-based methods are more efficient at registering than area-based methods when the images are very rich in geometrical details, as it is the case for remote sensing images. But they present, nevertheless, the problem
of being computationally more costly because the number of distinctive points to be calculated for these images is high. The authors proposed HSI-KAZE, a version of A- KAZE using the M-SURF descriptor that is especially adapted to hyperspectral images, as the spectral information is taken into account.
In this paper the implementation of the HSI-KAZE registration algorithm on programmable GPUs is explored as an attempt to reduce its computational cost. The GPU implementation focuses on reducing the cost of the two most costly steps of the algorithm, keypoint detection and
Keywords: hyperspectral data, image registration, KAZE features, remote sensing