鲁棒的掌纹图像特征提取方法  

A robust method to extract features from palm image

作　　者：竺乐庆[1,2] 陈文智[1] 幸锐[1] 张三元[1] 

机构地区：[1]浙江大学计算机科学与技术学院,浙江杭州310027 [2]浙江工商大学计算机与信息工程学院,浙江杭州310018

出　　处：《浙江大学学报（工学版）》2009年第8期1412-1418,共7页Journal of Zhejiang University：Engineering Science

基　　金：国家自然科学基金资助项目(60473106);国家"863"高技术研究发展计划资助项目(2007AA01Z311;2007AA04ZA5)

摘　　要：提出了一种基于低分辨率彩色图像的鲁棒的掌纹图像特征提取方法.采用均值平移算法对彩色图像帧中像素进行聚类,应用Ostu二值化方法分割出手掌,并提取出有效掌纹区域.采用KLT角点检测算法提取出有效掌纹区域内的特征点,给每个特征点赋予方向,并根据局部区域特征构造方向不变的特征向量,所有特征点及其特征向量的集合构成了掌纹图像特征.在识别时只须在两个特征点集之间查找匹配对应,并通过随机采样一致性检验最大一致集中内点个数是否大于自适应域值来确定两个手掌是否匹配.利用该算法对网络摄像头采集的手掌样本进行了实验测试,获得了较高的识别精度与性能.该算法对手掌的距离、方向、姿势没有特殊要求,是一种鲁棒高效的掌纹图像特征提取方法.A robust method was proposed to extract palmprint features from low resolution color images. Firstly, mean shift algorithm is used to cluster color blocks; secondly, Ostu binarization method is used for segmentation; thirdly, the feature points are extracted with KLT corner detector, to which an orientation is assigned according to the most fast varying gradient direction. The keypoint descriptor can be constructed relative to this orientation so that it＇s invariant to image rotation. The location, orientation and rotation invariant local descriptors of keypoints make up of the features of a palm image. In recogni- tion, testing palm features are matched to the template features for initial correspondences, then random sample consensus （RANSAC） is used to confirm the final matching. Plenty of experiments showed that palmprints could be recognized with high accuracy. The method has no limitation to palm rotation, very loose limitation to palm distance and pose, which further proves its robusticity.

关 键 词：掌纹识别 均值平移 OSTU 角点检测 随机采样一致性(RANSAC) 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]