计算机工程与应用 ›› 2023, Vol. 59 ›› Issue (2): 85-93.DOI: 10.3778/j.issn.1002-8331.2108-0349

• 理论与研发 • 上一篇    下一篇

自适应高效深度跨模态增量哈希检索算法

周坤,徐黎明,郑伯川,谢亦才   

  1. 1.西华师范大学 计算机学院,四川 南充 637009
    2.物联网感知与大数据分析南充市重点实验室,四川 南充 637009
    3.重庆邮电大学 计算机科学与技术学院,重庆 400065
  • 出版日期:2023-01-15 发布日期:2023-01-15

Adaptively Efficient Deep Cross-Modal Hash Retrieval Based on Incremental Learning

ZHOU Kun, XU Liming, ZHENG Bochuan, XIE Yicai   

  1. 1.School of Computer Science, China West Normal University, Nanchong, Sichuan 637009, China
    2.Internet of Things Perception and Big Data Analysis Key Laboratory of Nanchong, Nanchong, Sichuan 637009, China
    3.School of Computer Science and Technology, Chongqing University of Posts and Telecommunications, Chongqing 400065, China
  • Online:2023-01-15 Published:2023-01-15

摘要: 针对现阶段深度跨模态哈希检索算法无法较好地检索训练数据类别以外的数据及松弛哈希码离散化约束造成的次优解等问题,提出自适应深度跨模态增量哈希检索算法,保持训练数据的哈希码不变,直接学习新类别数据的哈希码。同时,将哈希码映射到潜在子空间中保持多模态数据之间的相似性和非相似性,并提出离散约束保持的跨模态优化算法来求解最优哈希码。此外,针对目前深度哈希算法缺乏有效的复杂度评估方法,提出基于神经网络神经元更新操作的复杂度分析方法,比较深度哈希算法的复杂度。公共数据集上的实验结果显示,所提算法的训练时间低于对比算法,同时检索精度高于对比算法。

关键词: 增量学习, 哈希编码, 语义保持, 潜在空间, 跨模态检索

Abstract: To address the problems that current deep learning-based cross-modal hash retrieval algorithms cannot retrieve new category data and sub-optimal solution caused by relaxing discretization constraint of hash codes, an adaptive deep incremental hashing(ADIH) retrieval algorithm is proposed to directly learn the hash codes of newly coming data meanwhile keeping the old trained data unchanged. In order to preserve the similarity and dissimilarity among multi-modal data, hash codes will be projected into latent semantic space where binary constrained discrete cross-modal hash algorithm is introduced to optimize hash code without using any relaxation. Besides, considering that there is currently no effective method which can be used to evaluate complexity of deep hashing methods, a novel method based on neuron updating operation is proposed to analyze the complexity. The experimental results on the public datasets show that the training time of the proposed algorithm is much lower than that of the comparison algorithms, and the retrieval accuracy is higher than that of the comparisons.

Key words: incremental learning, hash coding, semantic preservation, latent space, cross-modal retrieval