Author Login Editor-in-Chief Peer Review Editor Work Office Work

Computer Engineering ›› 2020, Vol. 46 ›› Issue (7): 260-267,276. doi: 10.19678/j.issn.1000-3428.0055259

• Graphics and Image Processing • Previous Articles     Next Articles

Deep Hash Learning Model Based on High-Order Statistical Information

GU Yan, ZHAO Chongyu, HUANG Ping   

  1. College of Physics and Optoelectronics, Taiyuan University of Technology, Jinzhong, Shanxi 030600, China
  • Received:2019-06-20 Revised:2019-07-22 Published:2019-08-06

基于高阶统计信息的深度哈希学习模型

顾岩, 赵崇宇, 黄平   

  1. 太原理工大学 物理与光电工程学院, 山西 晋中 030600
  • 作者简介:顾岩(1992-),男,硕士研究生,主研方向为深度学习、机器学习、数据分析与挖掘;赵崇宇,本科生;黄平(通信作者),教授。
  • 基金资助:
    山西省自然科学基金(201801D121020,201801D221132)。

Abstract: Deep hashing has been widely used in large-scale image retrieval for its advantage in retrieval efficiency and storage cost.To enhance the identification ability of hash code and improve the retrieval accuracy and efficiency,this paper proposes a deep hash learning model,BCI-DHH,based on high-order statistical information.The improved VGG-m model is employed to extract intra-layer auto-correlation features and inter-layer cross-correlation features from input images,and to generate a normalized high-order statistical vector.Then the weighting parameters are introduced to balance the number of positive and negative samples,and on this basis a contrastive loss function based on data balance is proposed.Then multi-level index hash blocks corresponding to dissimilar image pairs are differentiated to increase the hamming distance between the dissimilar image and the to-be-retrieved image,so as to optimize the compatibility of multi-level hash index.Experimental results on the benchmark datasets demonstrate that the proposed model outperforms BDH,DSH and other methods in terms of retrieval accuracy and efficiency.

Key words: deep hash, image retrieval, hash learning, higher-order statistics, contrastive loss, multi-level index

摘要: 深度哈希因其检索效率和存储代价上的优势而被广泛应用于大规模图像检索领域。为增强哈希编码的区分能力并提高检索准确率和效率,建立一种基于高阶统计信息的深度哈希学习模型BCI-DHH。采用改进的VGG-m模型分别提取输入图像基于层内的自相关特征和基于层间的互相关特征,并生成归一化的高阶统计向量。通过引入权重参数对训练样本中的正负样本数目进行平衡,提出一种基于数据平衡性的对比损失函数。在此基础上,对不相似图像对之间对应的多级索引哈希块进行差异化操作,增大不相似图像与其查询图像之间的汉明距离,优化多级哈希索引的兼容性。在基准数据集上的实验结果表明,该模型在检索准确率和效率方面优于BDH、DSH等方法。

关键词: 深度哈希, 图像检索, 哈希学习, 高阶统计, 对比损失, 多级索引

CLC Number: