图像尺度不变特征提取新方法

New Method for Scale-invariant Feature Extracting of Images

在文献[1]的基础上对尺度不变核进行深入研究,从理论上解决了初始函数选择的问题,构造了一族对尺度缩放具有不变性的核函数,能有效提取图像的尺度不变特征。核函数用于目标识别时,将图像发生尺度变换时数字化抽样插值所产生的误差引入到核参数的优化中,使得核函数能根据目标调整参数,达到最优的识别效果。实验结果表明,当用于二元目标的识别分类时,基于尺度不变核的方法能取得很好的识别效果。

This paper presents a new scale-invariant feature extracting method using kernels. The method is an extension of the framework of finding geometric invariant features proposed by Chen et al[ 1 ]. A deeper research of initial function is provided and a family of kernels invariant to scale is constructed. When used for target recognition, the error of sampling of images is considered into the Fisher like rule function, which provides an optimal chose of kernel parameters. It is shown that the new method is well suitable for classification of two classes＇ problem in our experiments.