高效并行递归高斯SIFT算法的实现

Efficient parallel recursive Gaussian SIFT algorithm

针对传统尺度不变特征变换scale invariant feature transform(SIFT)算法中计算复杂度高、实时性差的问题,提出一种基于多核处理器的数据级并行递归高斯-尺度不变特征变换(recursive Gaussian filter-scale invariant feature transform,RGFSIFT)算法。利用四阶递归高斯滤波逼近尺度不变特征变换算法中的线性高斯滤波,通过EDMA数据传输技术,将图像数据分割为多块,分配到多个DSP核并行处理。实验结果表明:并行递归高斯-尺度不变特征变换算法的特征点重复率比SIFT算法的高;在图像特征点个数小于或等于500的情况下,多核并行递归高斯-尺度不变特征变换算法的平均加速比为17.97倍。

A parallel RGF-SIFT algorithm is proposed to solve the problem of SIFT algorithm on high computational complexity and poor real-time,exploiting multi-core processor.Fourth-order recursive Gaussian filter is applied to approach linear Gaussian filtering of SIFT algorithm.Then image data are cut multi-block to allocate to multi-core for parallel processing through EDMA data transmission technology.The experimental results show parallel RGF-SIFT algorithm presents higher repetition rate.In the case of feature points less than or equal to five hundred,execution time of parallel RGF-SIFT algorithm accelerated ratio is 17.97 on average.