摘要
针对图像检索任务中部分监督学习部署困难,以及一般无监督学习没有利用监督信息导致检索性能劣化的问题,提出一种基于正态分布的距离保持哈希的无监督框架,使生成的哈希码保持图像的原始距离关系,在检索结果中尽可能保留相似的图像;距离保持哈希使用正态分布框架约束生成的连续码保持原始特征的距离关系,将图像的语义信息尽可能保留到哈希码中,并使用标准化欧氏距离进行特征向量的相似度计算,解决直接使用传统欧氏距离作为损失时神经网络梯度下降产生的网络波动问题。实验结果表明,与其他模型相比,距离保持哈希在大规模图像检索中检索性能更好。
In view of problems of difficult deployment of some supervised learning in image retrieval tasks,and poor retrieval performance due to lack of supervision information in generally unsupervised learning,an unsupervised framework of distance-keeping hashing based on normal distribution was proposed,which made generated hash codes maintain original distanc e relationship of images and retain similar images as far as possible in retrieval results.Distance-keeping hashing used continuous codes generated by using normal distribution framework constraints to keep the distance relationship of original features.Image semantic information was kept into hash codes as far as possible,and standardized Euclidean distance was used to calculate similarity of characteristic vectors.The problem of network fluctuation caused by gradient descent of neural network was solved when traditional Euclidean distance was directly used as a loss.The experimental results show that compared with other models,distance-keeping hashing has better performance in large-scale image retrieval.
作者
闫冰琦
田爱奎
吴楠楠
孙妍
王振
YAN Bingqi;TIAN Aikui;WU Nannan;SUN Yan;WANG Zhen(School of Computer Science and Technology,Shandong University of Technology,Zibo 255049,Shandong,China)
出处
《济南大学学报（自然科学版）》
CAS
北大核心
2022年第2期119-126,共8页
Journal of University of Jinan(Science and Technology)
基金
国家自然科学基金项目(61841602)
山东省自然科学基金项目(ZR2018PF005)。
关键词
图像检索
距离保持哈希
正态分布
语义信息
标准化欧氏距离
imag e retrieval
distance-keeping hashing
normal distribution
semantic i nformation
standardized Euclidean distance
