摘要
高分辨率数字视频图像数据量巨大,基于SIFT图像配准算法在CPU上实现时用时巨大。针对此,首先对配准算法中3个最耗时的部分:SIFT特征提取;SIFT特征匹配;RANSAC算法提纯匹配点对,求解变换模型参数。对此展开重点研究,研究其并行算法。然后基于CUDA并行快速实现高分辨率数字视频图像配准。实验结果表明:基于SIFT图像配准算法在CPU与CUDA上实现,在配准效果相近时,在CUDA上实现的处理速度比在CPU上实现的处理速度提高了100多倍,并且随着图像像素数的增加加速比有显著提高。
High-resolution digital video image contains a huge amount of data, and the （ scale invariant feature trans- form） SIFT based image registration algorithm achieved on CPU costs enormous amount of time. Aiming at this prob- lem, firstly we emphatically studied the parallel algorithms of the three most time-consuming parts in the image regis- tration algorithm： SIFT feature extraction; SIFT feature matching; and random sample consensus（RANSAC） algo- rithm for purifying matching point pair and solving transformation model parameters. Then, the high-resolution digital video image registration was achieved quickly and in parallel based on compute unified device architecture（CUDA）. The experiment results show that for the SIFT-based image registration algorithm implemented on CPU and CUDA, the processing speed for CUDA is over 100 times faster than that for CPU when the registration results are similar, and the acceleration ratio improves significantly while the number of the image pixels increases.
出处
《仪器仪表学报》
EI
CAS
CSCD
北大核心
2014年第2期380-386,共7页
Chinese Journal of Scientific Instrument
基金
航空科学基金(20135152049)
南京航空航天大学基本科研业务费专项科研项目(NN2012083
NS2010214
NP2011048)资助
关键词
图像配准
高分辨率
数字视频
CUDA
image registration
high-resolution
digital video
compute unified device architecture （CUDA）
