摘要
传统基于分类学习的监督哈希方法并不能完全满足哈希检索技术需求,但是线性判别分析却能够在一定程度上做到这一点.本文提出将线性判别分析作为深度网络的优化目标,以端到端训练的方式学习有效的哈希编码.但是,直接以上述目标训练神经网络就必须解决具有较高计算复杂度的特征值分解问题.在本文中,线性判别分析目标被转化为一个简单的最小均方问题,这种转化可以解决上述问题,同时可以利用成熟的优化方法优化网络.这种基于线性判别分析的深度网络拓展可以弥补传统判别分析在简单线性投影和特征学习上的劣势.本文在3个基准数据集上进行大量对比实验,相对于传统线性判别分析,本文所提方法在检索基准指标上有70%的提升,并超过大多数基于深度模型的哈希方法,这些实验结果证明了本文方法的有效性.
The conventional supervised hashing methods based on classification do not entirely meet the requirements of the hashing technique;on the other hand,the linear discriminant analysis(LDA)approach does satisfy such demands.In this paper,we propose to perform the LDA objective over deep networks to learn efficient hashing codes in a truly end-to-end fashion.However,a complicated eigenvalue decomposition within each mini-batch in every epoch has to be faced with when simply optimizing the deep network with respect to the LDA objective.Here,the LDA objective is transformed into a simple least-square problem,which naturally overcomes the intractable problems and can be easily solved by an off-the-shelf optimizer.Such deep extension can also overcome the weaknesses of LDA hashing involving the limited linear projection and feature learning.Numerous experiments were conducted on three benchmark datasets.The proposed deep LDA hashing approach exhibits an improvement of nearly 70 points for the CIFAR-10 dataset over the conventional strategy.Additionally,the proposed approach is found to be superior to several state-of-the-art methods for various metrics.
作者
胡迪
聂飞平
李学龙
Di HU;Feiping NIE;Xuelong LI(School of Artificial Intelligence,Optics and Electronics(iOPEN),Northwestern Polytechnical University,Xi’an 710072,China)
出处
《中国科学：信息科学》
CSCD
北大核心
2021年第2期279-293,共15页
Scientia Sinica(Informationis)
基金
科技部重点研发计划(批准号:2018YFB1107400)
国家自然科学基金(批准号:61871470,61772427)资助项目。
关键词
哈希技术
线性判别分析
最近邻检索
深度网络
量化技术
hashing technique
linear discriminant analysis
nearest-neighbor search
deep network
quantization technique
