GPU-Accelerated Algorithm to Compute Bessel-Fourier Moments

Abstract

Bessel-Fourier moments have been applied in image pattern reconstruction since their introduction in 2010. In this research, a scalable GPU-based algorithm is proposed to accelerate the computation of Bessel-Fourier moments of high orders while preserving accuracy. To analyze our new algorithm, image reconstructions from Bessel-Fourier moments of orders up to 1000 were tested on two systems. The experimental results prove the correctness and scalability of the algorithm. In addition, by investigating the precision-related performance, both 64-bit and 32-bit precisions were shown to provide the same level of computational accuracy for Bessel-Fourier moments of orders up to 1000. Nevertheless, reconstructions with 64-bit precision are computationally more costly. Furthermore, we applied filtering in Bessel-Fourier moments and Fourier Frequency domains and found that Bessel-Fourier moments share some similarities with the frequencies in Fourier Frequency domain, though more image power is distributed in the Bessel-Fourier moments of lower orders.