Image Geometric Correction Parallelization of the Multi-azimuth UV Imager Based on GPU

摘要

The Multi-Azimuth UV Imager (MAVI) is designed and built by Chinese scientific institute independently. The MAVI is comprised of the ultraviolet ring imager and the ultraviolet forward imager. One track of image of ultraviolet ring imager contains hundreds of standard image frames arranged along the track direction and possesses relatively higher data volume. To produce the standard level-one product rapidly, the entire procedure of geometric correction was decomposed into parameter parsing, data reading into memory, model of geometric correction, and data exporting to hard disk. Among these four steps of ultraviolet ring imager image processing, the construction and calculating of geometric correction model is the most time consuming. There exist several for loops in steps of geometric positioning and vector calculating. Simultaneously, the majority of C code of geometric positioning and vector calculating should be rewritten to adapt the CUDA GPU in this practice. Experimental results show that parallelized geometric correction could reach a speedup of 3.2 through image partition along the column direction on GPU. A better speedup result of URI image processing would be achieved through parallelizing all for loops in every step.