PARALLEL VERSION OF DETECTOR OF EXTREMAL KEY POINTS ON IMAGES

Abstract

The article presents a parallel version of the detector of extremal key points, which are used to describe, analyze and compare digital images by local descriptors. Local descriptors are determined in neighborhoods of the extremal key points. The orientation of the descriptors are found with aid of Histograms of Oriented Gradient. The specificity of the parallel architecture of NVIDIA graphics cards has been taken into account in the developed version, oriented to the implementation on CUDA. It accelerated the calculation of the extremal key points by several orders. Computation of the not oriented extremal key points for images of the FullHD-size on the budget graphics card takes 5–6 ms. The oriented extremal key points are computed within 11–12 ms.