尺度与颜色不变性图像特征描述

Scale and Color Invariant Image Feature Description

尺度不变特征变换是目前公认的鲁棒性最强的图像特征描述方法之一,在尺度不变性和几何不变性方面具有较好的特性,但该方法主要适用于灰度图像,对图像颜色的区分能力不强,因此,一些对象可能会因为颜色的不同而被错误的区分.另外,尺度不变特征变换对关键点局部范围内描述子主方向的依赖性非常强,直接决定了匹配的正确率,但是研究表明,主方向分配产生的误差仅有三分之二左右能控制在[-20。,+20。]范围内,因此部分特征会有三分之一的概率因为主方向分配的误差较大而不能正确匹配.针对以上两个问题,本文提出了一种具有颜色和尺度不变性的局部特征描述方法,颜色不变性通过将RGB图像转换到高斯颜色模型下实现,特征描述过程中不再分配主方向,而用局部相对方向,尺度不变性通过构建高斯金子塔实现.实验选取阿姆斯特丹数据集图像进行了测试,结果表明本文方法比传统尺度不变特征变换方法,在特征点的数目、分布均匀性以及匹配精度方面均有所提高.

Scale invariant feature transform is considered as one of the most robust image local feature description methods with satisfied scale invariance and geometry invariance.It is mainly designed for gray scale image with little discrimination to image color.Thus,some objects with different color may be misclassified.In addition,SIFT greatly relies on the allocation of descriptor canonical direction around the key point.But studies show that there about two thirds of the allocated canonical direction error is in the range of [-20.,＋20.],there about one third of local features may be mismatched for the over large direction error.To these two problems,this paper proposed a local feature descriptor with color and scale invariance,where color invariance is designed by converting the RGB image to Gaussian color model.The canonical direction is not allocated in the process of building descriptors but with local relative direction instead.And the scale invariance is designed by constructing Gaussian pyramid.Experiments were performed on the Amsterdam Library of Object Images dataset and Oxford dataset and the results show the improvement in key points count,distribution and matching precision than scale invariant feature transform.