
2019 年 1 月 Journal on Communications January 2019
2019021-1
第 40 卷第 1 期 通 信 学 报 Vol.40
No.1
仿视网膜采样的二进制描述子
袁庆升
1,3,4
,靳国庆
2
,张冬明
1,4
,包秀国
1,4
(1. 国家计算机网络应急技术处理协调中心,北京 100029;2. 中国科学院计算技术研究所,北京 100190;
3. 中国科学院大学网络空间安全学院,北京 100049;4. 中国科学院信息工程研究所,北京 100193)
摘 要:现有二进制描述子生成采用随机点对或均匀采样方式,顽健性弱、计算复杂。针对这一问题,提出了一
种模仿人眼视网膜特性的采样模式(RBS),首先通过设计采样密度、多尺度光滑、视野重叠等采样方法来模仿
视网膜神经节细胞层(ganglion cell layer),也称为视神经层,将光信号转换为视信息的方式,再通过对典型数据
学习来选择特征点对,最后使用区块均值代替单像素点计算点对比较值,生成顽健的紧致二进制描述子。在
Mikolajczyk 提出的数据集上进行了实验,实验结果表明,128 bit 的 RBS-128 相对于 512 bit 的 FREAK 和 BRISK
正确率分别提升 16.4%和 5.3%。
关键词:二进制描述子;仿视网膜采样;神经节细胞;比较点对
中图分类号:TP37
文献标识码:A
doi: 10.11959/j.issn.1000−436x.2019021
Retina-imitation sampling based binary descriptor
YUAN Qingsheng
1,3,4
, JIN Guoqing
2
, ZHANG Dongming
1,4
, BAO Xiuguo
1,4
1. National Computer Network Emergency Response Technical Team/Coordination Center of China, Beijing 100029, China
2. Institute of Computing Technology, Chinese Academy of Sciences, Beijing 100190,China
3. School of Cyber Security, University of Chinese Academy of Sciences, Beijing 100049, China
4. Institute of Information Engineering, Chinese Academy of Sciences, Beijing 100193,China
Abstract: The existing binary descriptors,generated from random or uniform point pairs sampling,suffer from low ro-
bustness and high computation. A novel sampling method, named RBS (retina-imitation based sampling), was proposed,
which combines different densities sampling, multi-scale smoothing and reception field overlapping to imitate the convert-
ing from light signal to vision of ganglion cells of human retina cells, and further selects most discriminative comparison
pairs based on learning on training data. Finally,compact binary descriptor was generated based on comparisons between the
neighbor mean instead of singe sampled point. The experimental results show the RBS-128 with 128 bit outperforms
FREAK and BRSIK with 512 bit about 16.4% and 5.3% in precision on the dataset provided by Mikolajczyk.
Key words: binary descriptor, retina-imitation based sampling, ganglion cell, comparison pairs
1 引言
随着互联网图像数量迅速增长,对海量图像进
行内容分析与检测的需求越来越大,基于内容的图
像检索技术成为研究热点。这些技术通常使用局部
特征作为图像内容的描述,这是由于局部特征具有
良好的区分性,对图像的多种变换,如遮挡、模糊、
噪声、剪切等,具有较高的稳定性。局部特征的提
取主要分为 2 个步骤:1) 特征点的提取,提取的信
息包括特征点位置、特征点主方向、特征点尺度信
收稿日期:2018−03−13;修回日期:2018−12−01
通信作者:张冬明,dmzhang@ict.ac.cn
基金项目:国家自然科学基金资助项目(No.61672495,No.61273247);国家重点研发专项基金资助项目(No.2016YFB0801203,
o.2016YFB0801200)
Foundation Items: The National Natural Science Foundation of China (No.61672495, No.61273247), The National Key Researc
and Development Program of China (No.2016YFB0801203, No.2016YFB0801200)