基于特征金字塔网络的余弦四元组哈希图像检索方法  

Deep cosine quadruplet hashing based on feature pyramid network for image retrieval

作　　者：盖枚岭 张辉辉 秦琦冰 GE Mei-ling;ZHANG Hui-hui;QIN Qi-bing(School of Computer Engineering,Weifang University,Weifang 261061,China;School of Information Science and Engineering,Ocean University of China,Qingdao 266000,China)

机构地区：[1]潍坊学院计算机工程学院,山东潍坊261061 [2]中国海洋大学信息科学与工程,山东青岛266000

出　　处：《计算机工程与设计》2024年第7期2127-2133,共7页Computer Engineering and Design

基　　金：山东省自然科学基金项目(ZR2021MC044、ZR2022QF046)。

摘　　要：为提高哈希图像检索的准确性,设计并提出一种基于特征金字塔网络的余弦四元组哈希图像检索方法,增强生成哈希编码的区分性。提出一种基于特征金字塔网络的特征提取器,提取到包含多层视觉信息和语义信息的图像特征描述符。设计基于余弦度量的四元组排序损失,使哈希码能够保持相似近邻关系;引入分类损失和二进制约束损失,使离散编码包含更多语义信息。实验结果表明,所提模型具有更好的检索性能。To improve the accuracy of hash image retrieval,the deep cosine quadruplet hashing based on feature pyramid network for image retrieval was proposed,which significantly enhanced the discriminative capability of learned binary codes.The feature extractor based on feature pyramid networks was to extract image feature descriptors containing multi-layer visual information and semantic information.The quadruplet ranking loss based on cosine metrics was designed to preserve the similar neighbor relationships between binary codes and training samples.The classification loss and binary code constraints were introduced,which enabled the generated compact codes to contain more semantic information.Experimental results show that the proposed model has better retrieval performance.

关 键 词：深度哈希 图像检索 特征金字塔 余弦度量 四元组损失 分类损失 二进制约束损失 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]