On the design of multiple key hashing files for concurrent orthogonal range retrieval between two disks

This paper concerns the following problem: given a set of multi-attribute records, a fixed number of buckets and a two-disk system, arrange the records into the buckets and then store the buckets between the disks in such a way that, over all possible orthogonal range queries (ORQs), the disk access concurrency is maximized. We shall adopt the multiple key hashing (MKH) method for arranging records into buckets and use the disk modulo (DM) allocation method for storing buckets onto disks. Since the DM allocation method has been shown to be superior to any other allocation methods for allocating an MKH file onto a two-disk system for answering ORQs, the real issue is knowing how to determine an optimal way for organizing the records into buckets based upon the MKH concept.

A performance formula that can be used to evaluate the average response time, over all possible ORQs, of an MKH file in a two-disk system using the DM allocation method is first presented. Based upon this formula, it is shown that our design problem is related to a notoriously difficult problem, namely the Prime Number Problem. Then a performance lower bound and an efficient algorithm for designing optimal MKH files in certain cases are presented. It is pointed out that in some cases the optimal MKH file for ORQs in a two-disk system using the DM allocation method is identical to the optimal MKH file for ORQs in a single-disk system and the optimal average response time in a two-disk system is slightly greater than one half of that in a single-disk system.     

Locality-preserving hash functions for general purpose parallel computation

Consider the problem of efficiently simulating the shared-memory parallel random access machine (PRAM) model on massively parallel architectures with physically distributed memory. To prevent network congestion and memory bank contention, it may be advantageous to hash the shared memory address space. The decision on whether or not to use hashing depends on (1) the communication latency in the network and (2) the locality of memory accesses in the algorithm.We relate this decision directly to algorithmic issues by studying the complexity of hashing in the Block PRAM model of Aggarwal, Chandra, and Snir, a shared-memory model of parallel computation which accounts for communication locality. For this model, we exhibit a universal family of hash functions having optimal locality. The complexity of applying these hash functions to the shared address space of the Block PRAM (i.e., by permuting data elements) is asymptotically equivalent to the complexity of performing a square matrix transpose, and this result is best possible for all pairwise independent universal hash families. These complexity bounds provide theoretical evidence that hashing and randomized routing need not destroy communication locality, addressing an open question of Valiant.This work was started when the author was a student at Oxford University, supported by a National Science Foundation Graduate Fellowship and a Rhodes Scholarship. Any opinions, findings, conclusions, or recommendations expressed in this publication are those of the author and do not necessarily reflect the views of the National Science Foundation or the Rhodes Trust.     

基于感知哈希的图像内容鉴别性能分析

感知哈希能有效地区分不同内容的图像,常用于检测盗版或重复的图像内容鉴别应用中.针对已有的对感知哈希的图像内容鉴别应用研究主要集中在算法设计上,缺少对其性能评价的理论方法的问题,提出从理论上评价内容鉴别性能的方法.该方法基于一类图像感知哈希算法的抗碰撞特性建立了内容鉴别的判定模型,并定义内容鉴别的性能公式,最后推导出统计学上的函数作为性能评价指标.实验结果表明,文中方法能很好地估计感知哈希在内容鉴别应用中的性能.     

稀疏差分网络和多监督哈希用于高效图像检索

针对基于深度哈希的图像检索中卷积神经网络（CNN）特征提取效率较低和特征相关性利用不充分的问题,提出一种融合稀疏差分网络和多监督哈希的新方法SDNMSH（sparse difference networks and multi-supervised hashing）,并将其用于高效图像检索。SDNMSH以成对的图像作为训练输入,通过精心设计的稀疏差分卷积神经网络和一个监督哈希函数来指导哈希码学习。稀疏差分卷积神经网络由稀疏差分卷积层和普通卷积层组成。稀疏差分卷积层能够快速提取丰富的特征信息,从而实现整个网络的高效特征提取。同时,为了更加充分地利用语义信息和特征的成对相关性,以促进网络提取的特征信息能够更加有效地转换为具有区分性的哈希码、进而实现SDNMSH的高效图像检索,采用一种多监督哈希（MSH）函数,并为此设计了一个目标函数。在MNIST、CIFAR-10和NUS-WIDE三个广泛使用的数据集上进行了大量的对比实验,实验结果表明,与其他先进的深度哈希方法相比,SDNMSH取得了较好的检索性能。     

一种具有平衡标签预测能力的在线哈希算法

针对传统离线哈希算法训练模型耗时、占用内存大和不易更新模型的问题,以及现实图像集的标签存在大量损失的现象,提出了一种能够平衡标签预测的在线哈希算法（BLPOH）。BLPOH通过标签预测模块生成预测标签,并融合残缺的真实标签,能够有效缓解因标签损失导致的模型性能下降。观察到标签存在分布不平衡现象,提出标签类别相似性平衡算法并应用于标签预测模块,提升标签预测的准确性。将旧数据的信息加入哈希函数的在线更新过程,提升模型对旧数据的兼容性。通过在两个广泛使用的数据集上进行实验,并和一些当前先进的算法进行对比,结果证实了BLPOH的优越性。     

基于DCT和感知哈希的纹理防伪标签鉴别算法

针对纹理防伪标签在光线不足、视力不佳的情况下难以鉴别的问题,提出了一种基于DCT、感知哈希和视觉特征向量的纹理防伪标签鉴别算法.首先通过DCT和感知哈希提取图像的特征向量;然后建立特征数据库,通过特征向量之间的归一化相关系数(NC)实现纹理防伪标签的自动鉴别.实验结果表明,该方法可以有效抵抗常规攻击和几何攻击,可实现自动鉴别技术.与现有的纹理防伪技术相比,该算法可以减小占据的数据库存储空间,具有更好的实用性.     

一种等量约束聚类的无监督蒸馏哈希图像检索方法

为了进一步降低无监督深度哈希检索任务中的伪标签噪声,提出了一种等量约束聚类的无监督蒸馏哈希图像检索方法。该方法主要分为两个阶段,在第一阶段中,主要对无标签图像进行软伪标签标注,用于第二阶段监督哈希特征学习,通过所提等量约束聚类算法,在软伪标签标注过程中可以有效降低伪标签中的噪声;在第二阶段中,主要对学生哈希网络进行训练,用于提取图像哈希特征。通过所提出的无监督蒸馏哈希方法,利用图像软伪标签指导哈希特征学习,进一步提高了哈希检索性能,实现了高效的无监督哈希图像检索。为了评估所提方法的有效性,在CIFAR-10、FLICKR25K和EuroSAT三个公开数据集上进行了实验,并与其他先进方法进行了比较。在CIFAR-10数据集上,与TBH方法相比,所提方法检索精度平均提高12.7%;在FLICKR25K数据集上,与DistillHash相比,所提方法检索精度平均提高1.0%;在EuroSAT数据集上,与ETE-GAN相比,所提方法检索精度平均提高16.9%。在三个公开数据集上进行的实验结果表明,所提方法能够实现高性能的无监督哈希检索,且对各类数据均有较好的适应性。     

基于深度学习的跨模态检索综述

由于多模态数据的快速增长,跨模态检索受到了研究者的广泛关注,其将一种模态的数据作为查询条件检索其他模态的数据,如用户可以用文本检索图像或/和视频。由于查询及其检索结果模态表征的差异,如何度量不同模态之间的相似性是跨模态检索的主要挑战。随着深度学习技术的推广及其在计算机视觉、自然语言处理等领域的显著成果,研究者提出了一系列以深度学习为基础的跨模态检索方法,极大缓解了不同模态间相似性度量的挑战,本文称之为深度跨模态检索。本文从以下角度综述有代表性的深度跨模态检索论文,基于所提供的跨模态信息将这些方法分为3类：基于跨模态数据间一一对应的、基于跨模态数据间相似度的以及基于跨模态数据语义标注的深度跨模态检索。一般来说,上述3类方法提供的跨模态信息呈现递增趋势,且提供学习的信息越多,跨模态检索性能越优。在上述不同类别下,涵盖了7类主流技术,即典型相关分析、一一对应关系保持、度量学习、似然分析、学习排序、语义预测以及对抗学习。不同类别下包含部分关键技术,本文将具体阐述其中有代表性的方法。同时对比提供不同跨模态数据信息下不同技术的区别,以阐述在提供了不同层次的跨模态数据信息下相关技术的关注点与使用异同。为评估不同的跨模态检索方法,总结了部分代表性的跨模态检索数据库。最后讨论了当前深度跨模态检索待解决的问题以及未来的研究方向。     

面向视觉搜索的空间局部敏感哈希方法

目的 视觉检索需要准确、高效地从大型图像或者视频数据集中检索出最相关的视觉内容,但是由于数据集中图像数据量大、特征维度高的特点,现有方法很难同时保证快速的检索速度和较好的检索效果。方法 对于面向图像视频数据的高维数据视觉检索任务,提出加权语义局部敏感哈希算法（weighted semantic locality-sensitive hashing, WSLSH）。该算法利用两层视觉词典对参考特征空间进行二次空间划分,在每个子空间里使用加权语义局部敏感哈希对特征进行精确索引。其次,设计动态变长哈希码,在保证检索性能的基础上减少哈希表数量。此外,针对局部敏感哈希（locality sensitive hashing, LSH）的随机不稳定性,在LSH函数中加入反映参考特征空间语义的统计性数据,设计了一个简单投影语义哈希函数以确保算法检索性能的稳定性。结果 在Holidays、Oxford5k和DataSetB数据集上的实验表明,WSLSH在DataSetB上取得最短平均检索时间0.034 25 s;在编码长度为64位的情况下,WSLSH算法在3个数据集上的平均精确度均值（mean average precision,mAP）分别提高了1.2%32.6%、1.7%19.1%和2.6%28.6%,与几种较新的无监督哈希方法相比有一定的优势。结论 通过进行二次空间划分、对参考特征的哈希索引次数进行加权、动态使用变长哈希码以及提出简单投影语义哈希函数来对LSH算法进行改进。由此提出的加权语义局部敏感哈希（WSLSH）算法相比现有工作有更快的检索速度,同时,在长编码的情况下,取得了更为优异的性能。     

Orientation Field Code Hashing: A Novel Method for Fast Palmprint Identification

For a large-scale palmprint identification system,it is necessary to speed up the identification process to reduce the response time and also to have a high rate of identification accuracy.In this paper,we propose a novel hashing-based technique called orientation field code hashing for fast palmprint identification.By investigating hashing-based algorithms,we first propose a double-orientation encoding method to eliminate the instability of orientation codes and make the orientation codes more reasonable.Secondly,we propose a window-based feature measurement for rapid searching of the target.We explore the influence of parameters related to hashing-based palmprint identification.We have carried out a number of experiments on the Hong Kong Poly U large-scale database and the CASIA palmprint database plus a synthetic database.The results show that on the Hong Kong Poly U large-scale database,the proposed method is about 1.5 times faster than the state-of-the-art ones,while achieves the comparable identification accuracy.On the CASIA database plus the synthetic database,the proposed method also achieves a better performance on identification speed.