针对尺度不变特征变换(SIFT)算法所提取图像特征点数量少、误匹率高的问题, 提出了一种基于高光谱图像的改进SIFT算法。首先, 依据传统SIFT算法中高斯金字塔的构造思想, 结合在不同波段下的高光谱图像具有相同宏观特征的特点, 首次用高光谱图像作为原始算法中经高斯变换产生的图像, 使得检测到的具有实际意义的特征点数量大幅增加; 其次, 传统SIFT算法以及大量的改进方法都只通过目标象元邻域范围内的像素信息来构造特征描述符, 而忽略了像素点的位置信息, 文中将目标象元的位置信息纳入了特征描述符, 在特征描述符的匹配阶段, 在利用邻域范围内的像素信息进行粗匹配之后, 利用特征描述符中的位置信息进行精细匹配。仿真实验结果表明: 在限定最优值与次优值之比的情况下, 采用高光谱图像构造高斯金字塔的方式能显著增加特征点的提取数量, 更多地挖掘出图像中的极值点; 在特征描述符中加入目标象元的位置信息作为特征点匹配第二阶段的判断依据, 正确匹配数量达到原方法的59倍以上, 极大提升了算法的匹配性能。

Aiming at the small number of feature points and high error rate in the traditional Scale Invariant Feature Transform(SIFT) algorithm, an improved SIFT algorithm was improved based on hyperspectral images. First, hyperspectral images were used as the images generated by Gaussian transformation based on the Gaussian pyramid construction in the traditional SIFT algorithmand the characteristics of hyperspectral images with the same macro-characteristics in different wavebands.This considerably increased the number of real significant feature points detected. Second, the traditional SIFT algorithm and several improved methods only construct the feature descriptor through the pixel information in the neighborhood of the target pixeland ignore the position information of the pixel. In this study, the position information of the target pixel was included in the feature descriptor. The pixel information in the neighborhood was first used for coarse matching, and the position information in the feature descriptor was subsequently used for fine matching. The simulation results showed that by limiting the ratio of the suboptimal value, the method of constructing Gaussian pyramid with hyperspectral images significantly increased the number of feature points extracted, and more extreme points in the image could be extracted.Furthermore, the position information of the target pixel was added to the feature descriptor as the judgment basis of the second stage of feature point matching. Consequently, the number of correct matching was at least 59 times that of the original method, which greatly improved the matching performance of the algorithm.