【作者】 张凯；

【导师】 李斌；

【作者基本信息】 中国科学技术大学 ， 信息与通信工程， 2018， 硕士

【摘要】 面向大规模图像数据的快速检索是计算机视觉领域的研究热点,有着重要的应用价值。随着多媒体和网络技术的发展,待查询数据库往往存储了千万量级的图片,图像检索系统面临着检索速度、检索结果排序、存储代价等多方面的挑战。其中,快速检索是大规模图像检索中最为迫切的需求。本文研究面向大规模图像数据的快速检索方法,主要研究工作和成果如下:一、提出了一种新的基于多测度的多倒排索引构建方法。目前主流的多倒排索引算法在构建多倒排索引时,对图像特征空间的划分是基于单一测度的,从而对图像样本之间的相似关系描述不够准确和充分。本文提出了多个倒排索引之间的互补性问题,通过引入多测度划分,有效地提升了多个倒排索引之间的互补性。本文提出了两种多测度的产生方法:基于简单随机采样的加权方法和基于超拉丁方采样的加权方法。理论分析表明,基于超拉丁方采样的加权方法定义的加权欧式距离,可以产生更加互补的多倒排索引。实验结果也表明,采用该方法产生的多倒排索引,可以获得更高的召回率。二、提出了一种基于果蝇算法的SAR图像快速检索方法。不同于倒排索引的向量量化,哈希是一种二值量化技术。果蝇算法是一种模仿果蝇嗅觉神经系统的哈希算法,也是目前最好的局部敏感哈希算法。本文修改了原有果蝇算法的量化方式,以更好地应用在SAR图像检索系统中。在典型SAR数据库上的实验结果验证了该方法的有效性和先进性。三、提出了一种基于多层损失函数的深度哈希方法(DSMH)。目前主流的深度哈希方法不能很好地处理多标签图像之间复杂的排序关系。本文设计了一种多层损失函数来区分多标签图像之间的相似排序关系。理论分析表明主流深度哈希方法DMSSPH以及本文提出的另一种DMSSPH的变体DSTLH,是DSMH的两个特例。在典型数据集上的实验结果验证了所提方法(DSMH)的有效性和先进性。更多还原

【Abstract】 Large scale image retrieval is a research hotspot in computer vision,and it also has great application value.With the rapid development of multimedia and Internet,the database often contains hundreds of millions of pictures.Large scale image retrieval faces a series of problems including retrieval speed,sorting results,storage cost and so on.In these problems,how to improve the retrieval speed has been the key problem of large scale image retrieval.The dissertation focuses on the problem of improving the retrieval speed in large scale image retrieval.The main research work and achievements can be listed as fol-lows:First,the dissertation proposes a new multiple inverted indexes method based on multiple metrics.In the mainstream multiple inverted indexes method,multiple in-verted indexes are produced independently or by random partition,which fails to ade-quately describe the similarity between the images.In this dissertation,the problem of the complementarity between multiple inverted indexes has been proposed and multiple metrics(weighted Euclidean distance)has been introduced to improve the complemen-tarity between multiple inverted indexes.Two methods of generating multiple metrics have been proposed:simple random sampling(SRS)weighted method and Latin square sampling(LHS)weighted method.The LHS weighted method can produce more differ-ent weighted Euclidean distance by theoretical analysis.Experiments on several public datasets show the effectiveness of the proposed method.Second,the dissertation proposed a SAR image retrieval method based on fly algo-rithm.Unlike vector quantization with inverted indexing method,hashing is a method based on binary quantization.The fly algorithm can generate hash codes for images effectively by mimicking the fruit fly olfactory circuit,which is the best local sensi-tive hashing method.The quantification method of the original fly algorithm has been changed to speed up the image retrieval.Experiments on several public SAR datasets show the effectiveness of the proposed method.Third,the dissertation proposed a new Deep Supervised Multilevel Hashing(DSMH)method.The mainstream deep hashing methods cannot deal with complex relations among multi-label images well.A multilevel loss function is elaborately designed to distinguish partial similarity between multi-label images.The state-of-the-art the method,DMSSPH,and its variant DSTLH are proved to be two special cases of the proposed method(DSMH).Experiments on several public datasets show the effective-ness of the proposed method.更多还原