改进的GPU-SIFT特征提取与匹配算法

Improved GPU-SIFT feature extraction and matching algorithm

针对SIFT算法得到的特征点数目太大、算法复杂耗时的问题,提出一种改进的SIFT特征提取与匹配算法并在GPU上进行了加速处理。通过分析算法的并行性,充分利用GPU多线程和存储器的优势对SIFT算法进行优化。在关键点精确定位过程中增加了第二次筛选,有效减少了特征点数量。发挥圆形具有旋转不变的优势,减少了算法的步骤同时描述符降到了64维。实验结果表明,该算法在保证匹配准确度的同时速度随图像复杂度的增强而提升,处理1600×1200图像时加速比可达2.3倍,提高了算法在实际应用中的实时性。

The algorithm of SITF is complicated and time-consuming,and the number of feature points obtained from it is too large.To solve above problems,a novel method of SIFT feature extraction and matching algorithms on the GPU method is proposed.Based on the analysis of algorithms＇ parallelism,the algorithm of SITF is optimized by making full use of advantages of GPU multithreading and memory.In order to effectively reduce the number of feature points,secondary screening is proposed in the process of key points precise positioning.The steps of the algorithm are reduced and descriptor drop to 64 dimensions by making full use of advantages of circular rotation invariant.The experimental results show that the proposed algorithms not only can ensure matching accuracy,but also improve calculation speed greatly.For example,to an image with the resolution of 1600Χ1200,nearly 2.3 times speed-up ratio can be acquired with our proposed method.Simultaneously,the real-time processing ability is improved in practical application.