基于光谱图像空间的改进SIFT特征提取与匹配

Improved SIFT Feature Extraction and Matching Based on Spectral Image Space

原始SIFT算法采用不同参数的高斯核取差,是对图像空间性质的一种测量方法.本文在光谱维度上取差,用光学系统在光谱维度上的差异作为图像空间性质的测量方法;传统SIFT方法及大量的改进方法只统计以特征点为中心的邻域范围内图像块的像素信息,文中将匹配过程分为2个步骤,首先利用邻域范围内的图像块像素信息进行粗匹配,然后选取排序后相似程度最高的4组匹配对作为基准匹配对,对特征点进行二次校验.仿真结果表明文中的设计方式显著增加了检测到的特征点数量,有效剔除了错误匹配.

As a method to measure the spatial properties of image, the Gaussian kernel with different parameters was used to get the difference in original SIFT algorithm, while the difference in the spectral dimension of optical system was used in the proposed method. Comparing with traditional sift method and a lot of improved methods, counting the image block pixel information only in the neighborhood around the feature points, the new method was arranged to divide the matching process into two steps. Firstly, the image block pixel information got from the neighborhood of the feature points was rough matched. And then four matching pairs with the highest similarity were selected as the benchmark matching pairs, and the feature points were checked twice. The simulation results show that the proposed method can significantly increase the number of detected feature points and effectively eliminate the error matching.