| 摘要: |
| 随着互联网空间中图像数据的爆发式增长和图像应用领域的拓宽, 大规模图像检索的需求与日俱增. 哈希学习为大规模图像检索提供显著的存储与检索效率, 并成为近年来一个研究热点. 现有哈希学习综述存在着时效性弱与技术路线不清晰的问题, 即多总结5–10年前的研究成果, 且较少总结哈希学习算法各组成部分间的关联关系. 鉴于此, 通过总结近20年公开发表的哈希学习文献, 对面向大规模图像检索的哈希学习进行系统的综述性研究. 首先, 介绍哈希学习的技术路线和哈希学习算法的主要组成部分, 包括损失函数、优化策略及样本外扩展映射. 其次, 将面向图像检索的哈希学习算法分为无监督哈希方法和监督哈希方法两类, 并分别梳理每类方法的研究现状和演化过程. 然后, 介绍哈希学习算法评估通用的图像数据集与评估指标, 并通过对比实验分析部分经典算法的性能. 最后, 结合哈希学习的局限性与新挑战对其发展前景进行阶段性总结与展望. |
| 关键词: 图像检索 大规模数据 近似最近邻检索 哈希学习 相似性保持 |
| DOI：10.13328/j.cnki.jos.007141 |
| 分类号: |
| 基金项目:国家自然科学基金(62176141, 62176139, 61876098, 62206160); 山东省自然科学基金青年项目(ZR2022QF082); 山东省自然科学基金重大基础研究项目(ZR2021ZD15); 山东省青年泰山项目(tsqn202103088); 山东省杰出青年基金(ZR2021JQ26); 山东建筑大学特聘教授专项 |
|
| Survey on Hash Learning for Large-scale Image Retrieval |
|
ZHANG Xue-Ning1, LIU Xing-Bo2, SONG Jing-Kuan3, NIE Xiu-Shan2, WANG Shao-Hua2, YIN Yi-Long1
|
|
1.School of Software, Shandong University, Jinan 250101, China;2.School of Computer Science and Technology, Shandong Jianzhu University, Jinan 250101, China;3.School of Computer Science and Engineering, University of Electronic Science and Technology of China, Chengdu 611731, China
|
| Abstract: |
| As image data grows explosively on the Internet and image application fields widen, the demand for large-scale image retrieval is increasing greatly. Hash learning provides significant storage and retrieval efficiency for large-scale image retrieval and has attracted intensive research interest in recent years. Existing surveys on hash learning are confronted with the problems of weak timeliness and unclear technical routes. Specifically, they mainly conclude the hashing methods proposed five to ten years ago, and few of them conclude the relationship between the components of hashing methods. In view of this, this study makes a comprehensive survey on hash learning for large-scale image retrieval by reviewing the hash learning literature published in the past twenty years. First, the technical route of hash learning and the key components of hashing methods are summarized, including loss function, optimization strategy, and out-of-sample extension. Second, hashing methods for image retrieval are classified into two categories: unsupervised hashing methods and supervised ones. For each category of hashing methods, the research status and evolvement process are analyzed. Third, several image benchmarks and evaluation metrics are introduced, and the performance of some representative hashing methods is analyzed through comparative experiments. Finally, the future research directions of hash learning are summarized considering its limitations and new challenges. |
| Key words: image retrieval large-scale data approximate nearest neighbor search hash learning similarity preserving |
