A Novel Parallelized Feature Extraction in Grouped Scale Space Based on Graphic Processing Units

Abstract

Feature extraction algorithms in nonlinear scale space (such as KAZE and A-KAZE) have much better robustness than those in Gaussian scale space (such as SIFT and SURF). However, the former is very time-consuming. Although graphic processing units (GPUs) have been used to accelerate the feature extraction in Gaussian scale space, it is still a big challenge to do this in nonlinear scale space by using GPUs. In this paper, a novel GPU-based feature extraction approach in nonlinear scale space is presented by introducing a new idea of grouped scale space. By decoupling octaves, different groups can be processed in parallel, which can only be done sequentially in Gaussian scale space. A data-package method is also presented to combine these images with different sizes in different groups to eliminate the load imbalance. Moreover, this kind of grouped scale space can even improve the robustness of feature extraction. Experimental results show that the proposed approach can achieve much faster speed than existing state-of-the-art works, even those with lower robustness.