基于改进AKAZE算法的图像特征匹配方法

Image feature matching method based on improved AKAZE algorithm

由于AKAZE算法中局部二进制描述符对图像尺度变化、模糊变化不敏感,导致特征点提取不均匀、特征匹配正确率低,提出一种改进AKAZE与边缘化采样一致方法(marginalizing sample consensus,MAGSAC)结合的特征匹配算法。该算法采用FREAK(fast retina keypoint)描述符描述特征点,对采样点计算梯度确定特征点主方向,使用MAGSAC方法剔除错误匹配点对。实验结果表明:在图像发生尺度与旋转变化时,改进算法的匹配精度比传统AKAZE算法高6.98%,尺度变化下特征点提取平均耗时较传统AKAZE算法减少0.084 ms,图像模糊变化时匹配精度比传统AKAZE算法高8.57%,特征点提取平均耗时较传统AKAZE算法减少0.05 ms。

The local binary descriptors in AKAZE algorithm are not sensitive to image scale change and fuzzy change,which leads to uneven feature point extraction and low accuracy of feature matching.An improved feature matching algorithm combining AKAZE and MAGSAC was thus proposed.In this algorithm,the FREAK descriptor was used to describe the feature points,and the main direction of the feature points was determined by calculating the gradient of the sample points.The MAGSAC method was used to eliminate the wrong match points.The experimental results showed that the matching accuracy of the improved algorithm was 6.98%higher than that of the traditional AKAZE algorithm when the image changed in scale and rotation,and the average time of feature point extraction was 0.084 ms less than that of the traditional AKAZE algorithm when the image changed in scale;the matching accuracy was 8.57%higher than that of the traditional AKAZE algorithm when the image changed in blur,and the average time of feature point extraction was 0.05 ms less than that of the traditional AKAZE algorithm.