基于标签对的深度哈希学习

发布时间：2018-03-19 19:04

本文选题：哈希学习　切入点：深度学习　出处：《南京大学》2017年硕士论文　论文类型：学位论文

【摘要】：在图像检索等真实应用场景中,随着数据爆发式增长,近似最近邻搜索(approximate nearest neighbor search,ANNS)最近几年已经变成了一个热门研究课题。现有的近似最近邻搜索技术中,哈希学习凭借着检索速度快和存储成本低的优点,已经成为图像检索领域最受欢迎和有效的技术之一。但是现有的大部分哈希学习方法都是基于手工提取特征,手工提取特征不一定适用于哈希编码学习。最近一些深度哈希学习被提出,可以同时进行特征学习和哈希编码学习。然而这些深度哈希学习方法大部分都是基于三元组标签的。基于标签对的图像检索是哈希学习另一个常见应用场景,在这个场景下目前还没有深度哈希学习方法能够同时进行特征学习和哈希编码学习。在本文中,我们提出一种新颖的深度哈希学习方法,叫做基于标签对的深度哈希学习方法(deep pairwise-supervised hashing,DPSH),可以在一个完整框架中同时进行特征学习和哈希编码学习。在真实数据集上的实验表明,我们提出的DPSH方法在图像检索任务上可以取得比其他经典哈希学习方法更好的检索准确率。尽管我们提出的DPSH方法已经在图像检索任务上取得了令人满意的结果,但是对于大规模数据集而言,基于单个GPU的实现比较费时。所以我们提出了一种多GPU并行DPSH方法,在不影响检索准确率的前提下明显降低了训练时间。
[Abstract]:In real-world applications such as image retrieval, as data explodes, approximate nearest neighbor search has become a hot research topic in recent years. Hash learning has become one of the most popular and effective techniques in image retrieval due to its advantages of fast retrieval speed and low storage cost. However, most of the existing hash learning methods are based on manual feature extraction. Manual feature extraction is not necessarily applicable to hashing coding learning. Recently, some deep hashing learning has been proposed. Both feature learning and hashing coding can be performed simultaneously. However, most of these deep hash learning methods are based on triple tags. Image retrieval based on tag pairs is another common application scenario for hashing learning. In this scenario, there is no deep hash learning method that can perform both feature learning and hash coding learning. In this paper, we propose a novel depth hash learning method. The deep hashing learning method called tag pair based deep pairwise-supervised hashingDPS HHN can be used in both feature learning and hashing coding learning in a complete framework. Experiments on real data sets show that, Our proposed DPSH method can achieve better retrieval accuracy than other classical hash learning methods in image retrieval tasks, although our proposed DPSH method has achieved satisfactory results in image retrieval tasks. However, for large scale data sets, the implementation based on a single GPU is time-consuming, so we propose a multi-#en1# parallel DPSH method, which can significantly reduce the training time without affecting the retrieval accuracy.
【学位授予单位】：南京大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】