Fast low-level multi-scale feature extraction for hexagonal images
Sonya Coleman, Bryan W. Scotney, Bryan Gardiner · 2017
Inspired by the human vision system and its capability to process in real-time, an efficient framework for low-level feature extraction on hexagonal pixel-based images is presented. This is achieved by utilizing the spiral architecture addressing scheme to simulate eye-tremor along with the convolution of non-overlapping gradient masks. Using sparse spiral convolution and the development of cluster operators, we obtain a set of output image responses “a-trous” that is subsequently collated into a consolidated output response; it is also demonstrated that this framework can be extended to feature extraction at different scales. We show that the proposed framework is considerably faster than using conventional spiral convolution or the use of look-up tables for direct access to hexagonal pixel neighbourhood addresses.