多模态鲁棒的局部特征描述符  

Multimodality robust local feature descriptors

作　　者：赵春阳[1,2,3] 赵怀慈[1,2] 

机构地区：[1]中国科学院沈阳自动化研究所,辽宁沈阳110016 [2]中国科学院光电信息处理重点实验室,辽宁沈阳110016 [3]中国科学院大学,北京100049

出　　处：《光学精密工程》2015年第5期1474-1483,共10页Optics and Precision Engineering

基　　金：国家973重点基础研究发展计划资助项目;中国科学院光电信息处理重点实验室开放基金资助项目(No.OEIP-O-201203)

摘　　要：针对基于灰度的局部特征匹配方法对图像对比度变化敏感,导致在多模态图像配准应用中性能大幅下降的问题,提出了一种多模态鲁棒的局部特征描述符和匹配方法。首先,基于对比度变化不敏感的相位一致性和局部方向信息,提出一种多模态鲁棒的角点和线段特征提取方法,在对比度差异显著的多模态图像之间提取较多的共性角点和线段特征;然后,以角点为中心选择48个均匀分布的圆形特征子区域,利用角点与特征子区域内线段的距离和线段长度信息,构建96维的特征向量;最后,将归一化相关函数作为匹配测度函数进行特征匹配,并采用基于位置约束的随机抽样一致（RANSAC）方法进行匹配提纯。实验表明,本文提出的多模态匹配方法匹配正确率和重复率分别高达80%和13%,分别为对称-尺度不变特征变换算法（S-SIFT）、多模态-快速鲁棒特征算法（MM-SURF）等基于灰度方法的2~4倍和4~7倍,显著优于同类方法。The intensity-based local feature matching methods are sensitive to image contrast variations,so the performance declines significantly when they are applied in multimodal image registration.To solve the above problem,a multimodality robust local feature descriptor was proposed and the corresponding feature matching method was developed.Firstly,an extraction method for the multimodality robust corner and line segment was proposed based on the phase congruency and local direction information insensitive to contrast variants.Compared with intensitybased method,more equivalent corners and line segments were extracted between multimodal imageswith more contrast differences.Then,the feature region containing of 48 circular sub-regions was selected by using the corner for a center and the 96 dimensional feature vectors were generated by using the distance values of corners and the length values of line segments located in feature subregions.Finally,the feature matching method based on normalized correlation function was proposed and the location constraint-based RANdom SAmple Consensus（RANSAC）algorithm was used to remove false matching point pairs. The experimental results indicate that the precision and repeatability on multimodal image matching of the proposed method reach 80%and 13%respectively.As compared with the other intensity-based image matching methods,the precision and repeatability of proposed method are 2-4times and 4-7times respectively those of Symmetric-Scale Invariable Feature Transformation（S-SIFT）and Multimodal-Speeded-up Robust Features（MM-SURF）.It concludes that the proposed method outperforms many state-of-the-art methods significantly.

关 键 词：图像配准 多模态配准 多模态鲁棒特征 相位一致性 局部方向 归一化相关 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]