Deep Multi-label Hashing for Large-Scale Visual Search Based on Semantic Graph
Deep Multi-label Hashing for Large-Scale Visual Search Based on Semantic Graph
复制标题
DOI：
10.1007/978-3-319-63579-8_14
复制
发表时间：
2017-07
影响因子：
2.2
通讯作者：
Chunlin Zhong;Yi Yu-;Suhua Tang;S. Satoh;Kai Xing
中科院分区：
文献类型：
--
作者：
Chunlin Zhong;Yi Yu-;Suhua Tang;S. Satoh;Kai Xing
随着时间的推移，大量的图像被聚合，因为许多人将他们喜欢的图像上传到各种社交网站，如Flickr，并与朋友分享。因此，从大型图像数据库中进行视觉搜索变得越来越重要。哈希是一种高效的大规模视觉内容搜索技术，基于学习的哈希方法由于近年来深度学习的进展而取得了巨大的成功。然而，现有的深度哈希方法大多集中在单标签图像上，哈希码不能很好地保持图像的语义相似性。本文提出了一种基于语义图的深度多标签哈希（deep multi-label hash， DMLH）框架，该框架由三个关键组成部分组成：(i)对语义相似的图像标签进行分类，使相似的标签在同一聚类中。这有助于为哈希学习提供准确的基础真理。（ii）训练深度模型，同时生成图像的哈希码和特征向量，在此基础上通过哈希表组织多标签图像数据库。该模型在提高检索速度的同时保持了图像间的语义相似度。（iii）结合基于哈希码的粗搜索和基于特征向量的精细图像排序，提供高效准确的检索。在几个大型图像数据集上进行的大量实验证实，所提出的DMLH方法优于最先进的有监督和无监督图像检索方法，平均精度的增益范围为6.25%至38.9%。
Huge volumes of images are aggregated over time because many people upload their favorite images to various social websites such as Flickr and share them with their friends. Accordingly, visual search from large scale image databases is getting more and more important. Hashing is an efficient technique to large-scale visual content search, and learning-based hashing approaches have achieved great success due to recent advancements of deep learning. However, most existing deep hashing methods focus on single label images, where hash codes cannot well preserve semantic similarity of images. In this paper, we propose a novel framework, deep multi-label hashing (DMLH) based on a semantic graph, which consists of three key components: (i) Image labels, semantically similar in terms of co-occurrence relationship, are classified in such a way that similar labels are in the same cluster. This helps to provide accurate ground truth for hash learning. (ii) A deep model is trained to simultaneously generate hash code and feature vector of images, based on which multi-label image databases are organized by hash tables. This model has excellent capability in improving retrieval speed meanwhile preserving semantic similarity among images. (iii) A combination of hash code based coarse search and feature vector based fine image ranking is used to provide an efficient and accurate retrieval. Extensive experiments over several large image datasets confirm that the proposed DMLH method outperforms state-of-the-art supervised and unsupervised image retrieval approaches, with a gain ranging from 6.25% to 38.9% in terms of mean average precision.