Fast Zernike Moment Computation Using PyTorch in a Multiple-GPU Environment

This paper addresses the high computational cost of Zernike moment calculation by introducing a fast computation method leveraging a multi-GPU environment. Using a unit circle inscribed image transformation, the proposed method reduces the moment calculation time by over 36% compared to traditional methods. Experiments conducted with the PyTorch framework on a multi-GPU setup (four 12 GB GPUs) demonstrated that the optimized approach significantly accelerates computation while maintaining accuracy. Specifically, the moment calculation time at order 30 with a 6 K unit circle grid resolution was reduced from 478.0 s on a CPU to just 19.4 s on the GPU. Additionally, the proposed Zernike image pyramid cache eliminates recalculation overhead, further improving efficiency. The results validate the feasibility of real-time Zernike moment computation and high-accuracy image restoration, with a PSNR of approximately 100 at order 30.