Design and implementation on multi-threaded CUDA-optimized framework for division-of-aperture full-Stokes real-time imaging polarimeter

Real-time polarimetric imaging with high resolution faces a challenge of heavy computation burden due to the complexities of image calibration and processing. Traditional CPU-based implementation framework of it encounters prohibitive bottlenecks, with processing times exceeding 200 ms per frame, fundamentally limiting application potentials. Additionally, most existing polarization cameras are limited to the linear polarization detection without the circular one. To address these issues, we propose and design a novel multi-threaded polarization image processing framework to simultaneously acquire three linear polarization states and one circular one for a self-developed full-Stokes division-of-aperture polarization (DoAP) camera. Our key innovation is a parallel computing architecture that utilizes GPU acceleration to enable real-time processing on consumer-grade laptop, achieving a stable output of 44 frames per second for polarization parameters images. Experimental results show that, as compared to the CPU-based framework, the processing speed of this CUDA-optimized GPU-based framework improves over 10 times, leading the per-frame time less than 23 ms. This work offers a practical solution for high-resolution real-time full-Stokes polarimetric imaging.