计算机应用 ›› 2021, Vol. 41 ›› Issue (7): 1902-1907.DOI: 10.11772/j.issn.1001-9081.2020091472

所属专题: 人工智能

• 人工智能 • 上一篇    下一篇

基于相关度距离的无监督并行哈希图像检索

杨粟1,2, 欧阳智1, 杜逆索1,2   

  1. 1. 贵州省公共大数据重点实验室(贵州大学), 贵阳 550025;
    2. 贵州大学 计算机科学与技术学院, 贵阳 550025
  • 收稿日期:2020-09-21 修回日期:2021-01-12 出版日期:2021-07-10 发布日期:2021-01-26
  • 通讯作者: 欧阳智
  • 作者简介:杨粟(1996-),女(侗族),贵州铜仁人,硕士研究生,主要研究方向:图像检索;欧阳智(1987-),男,四川安岳人,副教授,博士,主要研究方向:机器学习、大数据治理;杜逆索(1986-),男,贵州六盘水人,副教授,博士,CCF会员,主要研究方向:仿真模拟、数据科学。
  • 基金资助:
    贵州省科学技术厅重大科技计划项目(黔科合重大专项字[2018]3002)。

Unsupervised parallel hash image retrieval based on correlation distance

YANG Su1,2, OUYANG Zhi1, DU Nisuo1,2   

  1. 1. Guizhou Provincial Key Laboratory of Public Big Data(Guizhou University), Guiyang Guizhou 550025, China;
    2. College of Computer Science and Technology, Guizhou University, Guiyang Guizhou 550025, China
  • Received:2020-09-21 Revised:2021-01-12 Online:2021-07-10 Published:2021-01-26
  • Supported by:
    This work is partially supported by Major Scientific and Technological Program of Department of Science and Technology of Guizhou Province ([2018]3002).

摘要: 针对传统无监督哈希图像检索模型中存在图像数据之间的语义信息学习不足,以及哈希编码长度每换一次模型就需重新训练的问题,提出一种用于大规模图像数据集检索的无监督搜索框架——基于相关度距离的无监督并行哈希图像检索模型。首先,使用卷积神经网络(CNN)学习图像的高维特征连续变量;然后,使用相关度距离衡量特征变量构建伪标签矩阵,并将哈希函数与深度学习相结合;最后,在哈希码生成时使用并行方式逐步逼近原始视觉特征,达到一次训练生成多长度哈希码的目的。实验结果表明,该模型在FLICKR25K数据集上对16 bit、32 bit、48 bit和64 bit的4种不同哈希码的平均精度均值(mAP)分别为0.726、0.736、0.738和0.738,与SSDH模型相比分别提升了9.4、8.2、6.2、7.3个百分点;而在训练时间方面,该模型与SSDH模型相比减少6.6 h。所提模型在大规模图像检索时能够有效缩短训练时间、提升检索精度。

关键词: 图像检索, 卷积神经网络, 哈希算法, 无监督, 相关度距离

Abstract: To address the problems of insufficient learning of semantic information between image data and the need to retrain the model every time when the hash code length is changed in traditional unsupervised hash image retrieval model, an unsupervised search framework for large-scale image dataset retrieval, the unsupervised parallel hash image retrieval model based on correlation distance, was proposed. First, the Convolutional Neural Network (CNN) was used to learn the high-dimensional feature continuous variables of the image. Second, the pseudo-label matrix was constructed by using the correlation distance measure feature variables, and the hash function was combined with deep learning. Finally, the parallel method was used to gradually approximate the original visual characteristics during the hash code generation, realizing the purpose of generating the multi-length hash codes in one training. Experimental results show that the mean Average Precisions (mAPs) of the proposed model for four of 16 bit, 32 bit, 48 bit and 64 bits hash codes on FLICKR25K dataset are 0.726, 0.736, 0.738, 0.738,respectively, which are 9.4, 8.2, 6.2, 7.3 percentage points higher than those of Semantic Structure-based Unsupervised Deep Hashing (SSDH) model, respectively; and compared with SSDH model, the training time of the proposed model is reduced by 6.6 hours. It can be seen that the proposed model can effectively shorten the training time and improve the retrieval accuracy in large-scale image retrieval.

Key words: image retrieval, Convolutional Neural Network (CNN), hash algorithm, unsupervised, correlation distance

中图分类号: