Quasi-random scale space approach to robust keypoint extraction in high-noise environments

A novel multi-scale approach is presented for the purpose of robust keypoint extraction in high-noise environments. A multi-scale representation of the noisy scene is computed using quasi-random scale space theory. A gradient second-order moment analysis is employed at each quasi-random scale to identify initial keypoint candidates. Final keypoints and their characteristic scales are selected based on the local Hessian trace extrema over all quasirandom scales. The proposed keypoint extraction method is designed to reduce noise sensitivity by taking advantage of the structural localization and noise robustness gained through the use of quasi-random scale space theory. Experimental results using scenes under different high noise conditions, as well as real synthetic aperture sonar imagery, show the effectiveness of the proposed method for noise robust keypoint extraction when compared to existing keypoint extraction techniques.