改进的SIFT邻域投票图像匹配算法

Improved SIFT neighborhood voting image matching algorithm

针对传统图像匹配算法在几何差异场景下匹配精度低的问题,提出一种改进SIFT特征描述符和邻域投票相结合的图像匹配算法。使用8个邻域像素的平均值代替原始极值点,通过SIFT提取图像中的特征点,利用Sobel算子计算特征点的梯度幅度和方向,结合8个仿射形式的同心圆邻域生成64维描述符,根据欧氏距离确定初始匹配点,采用邻域投票的方法剔除错误的匹配点,实现图像的精确匹配。实验结果表明,该算法在显著提高匹配精度的同时缩短了匹配时间,对复杂场景的匹配性能明显提升。

Aiming at the low matching accuracy of traditional image matching algorithm in geometric difference scene,an improved image matching algorithm combined with SIFT feature descriptor and neighborhood voting was proposed.The average of 8 adjacent pixels was used to substitute the original extreme point,and the feature points in the image were extracted using SIFT.The gradient magnitude and direction of the feature points were calculated using Sobel operator,and the 64-dimensional descriptor was generated by combining eight affine concentric circle neighborhoods,and the Euclidean distance was used to determine the initial matching points.The neighborhood voting method was used to eliminate the wrong matching points,and the accurate matching of images was realized.The final experimental results show that the proposed algorithm not only improves the matching accuracy significantly,but shortens the matching time,and improves the matching performance of complex scenes significantly.