针对SIFT、SURF等局部不变特征在大尺寸图像上匹配时过于耗时的问题, 将FREAK算子应用于图像匹配中, 并提出一种多线程并行加速方法。首先介绍FREAK描述子的特征点的检测、特征描述向量的生成和特征向量的匹配的过程, 并分析其优势。其次提出并行处理的2种思路: 一是对待匹配图像进行有重叠的分块, 对于每一块子图像, 开辟新的线程分别进行处理; 二是对匹配过程的3个步骤, 采用流水线技术进行并行处理, 每检测出一个特征点, 随即提取出该点的特征向量, 然后和模板图像的特征向量集进行匹配。改写SIFT、SURF和FREAK算法进行实验验证, 结果证明FREAK计算过程比SIFT和SURF快得多, 而并行方法可以在保证匹配精度的同时明显缩短匹配时间。

Taking into account the problem that local invariant feature algorithms such as SIFT,SURF take too much time in large images registration,a fast descriptor named FREAK is introduced,and a speeding up method using multi-thread technology is proposed. The process of feature point detection, descriptive characteristics vector generation and feature vector matching of FREAK are introduced and their advantages are analyzed. Then,two ideas of parallel processing are proposed.One is to divide large image into overlapping sub-blocks, and for each sub-image, a new thread is opened up. The other is to use pipelining technology for the three step of registration. Once a feature point is detected, its feature vector is extracted immediately, and soon it matches with the feature vector set of the template image. Rewriting SIFT,SURF and FREAK for experiment,the experimental result indicates that FREAK is faster than the other two algorithms,and our parallel method can greatly economize matching time as well as guarantee the accuracy.