摘要
针对高维特征图像检索中的精度和速度挑战,提出了一种结合深度哈希技术和VP-Tree索引的快速图像检索方法.该方法首先设计了一个轻量级的深度卷积编码网络,并在网络中引入了卷积块注意力模块和空间金字塔池化技术,以增强特征提取能力;然后通过该网络模型将图像数据集中每幅图像的高维特征转化为二进制哈希编码,并与其对应的图像编号组成一个哈希表;接着使用所有图像的哈希编码来构建一个VP-Tree,在执行图像检索时将使用待查询图像的哈希编码从VP-Tree中快速找到与其距离最近的节点;最后根据这些节点的哈希值从哈希表中取出对应的结果图像.实验结果表明,所提方法在保持高检索精度的同时显著提升了检索速度(在MNIST、FASHION-MNIST和CIFAR-10上的检索速度分别提高了24.17、8.61和4.01倍).
This paper proposed a fast image retrieval method that combines deep hashing technology and VP-Tree indexing to address the precision and speed challenges in high-dimensional feature image retrieval.The method first designed a lightweight deep convolutional encoding network which introduced convolutional block attention modules and spatial pyramid pooling techniques to enhance feature extraction capabilities.Then,through this network model,the high-dimensional features of each image in the image dataset were transformed into binary hash codes,which were combined with their corresponding image IDs to form a hash table.Subsequently,a VP-Tree was constructed using the hash codes of all images.During image retrieval,the hash code of the query image was used to quickly find the nearest nodes in the VP-Tree.Finally,the corresponding result images were retrieved from the hash table based on the hash values of these nodes.Experimental results showed that the proposed method significantly improved retrieval speed while maintaining high retrieval accuracy(the retrieval speed on MNIST,FASHION-MNIST,and CIFAR-10 was increased by 24.17,8.61,and 4.01 times,respectively).
作者
吴宗胜
李红
薛茹
WU Zong-sheng;LI Hong;XUE Ru(School of Computer,Xianyang Normal University,Xianyang 712000,China;School of Information Engineering,Xizang Minzu University,Xianyang 712000,China)
出处
《西南民族大学学报(自然科学版)》
CAS
2024年第5期544-553,共10页
Journal of Southwest Minzu University(Natural Science Edition)
基金
国家自然科学基金资助项目(62073218)
陕西省科技厅自然科学基础研究计划面上项目(2023-JC-YB-524)。