查询结果:   郁延珍.基于深度多监督哈希的快速图像检索[J].计算机应用与软件,2019,36(11):229 - 234.
中文标题
基于深度多监督哈希的快速图像检索
发表栏目
图像处理与应用
摘要点击数
52
英文标题
FAST IMAGE RETRIEVAL BASED ON DEEP MULTI-SUPERVISED HASHING
作 者
郁延珍 Yu Yanzhen
作者单位
复旦大学计算机科学技术学院 上海 201203     
英文单位
School of Computer Science, Fudan University, Shanghai 201203, China     
关键词
图像检索 深度多监督哈希 卷积神经网络
Keywords
Image retrieval Deep multi-supervised hashing Convolutional neural network
基金项目
国家重点研发计划项目(2016YFC0801003);上海市科委科研计划项目(17511108504)
作者资料
郁延珍,硕士,主研领域:多媒体信息处理与检索。 。
文章摘要
由于较低的检索时间和空间复杂度,哈希方法被广泛应用于大规模图像检索领域。提出深度多监督哈希(Deep Multi-Supervised Hashing,DMSH)方法来学习具有高度判别能力和紧凑的哈希编码,并进行有效的图像检索。设计一个新的卷积神经网络结构来产生相似性保留的哈希编码,用一个识别信号来增加类间距离,用一个验证信号来降低类间距离。同时,通过正则化的方式降低网络输出和二进制哈希编码之间的损失并使二进制哈希值在每一维上均匀分布使网络输出更接近离散的哈希值。在两个数据集上的实验证明了该方法能够快速编码任意新的图像并取得先进的检索结果。
Abstract
Hashing methods have been widely applied in the field of large-scale image retrieval due to their low retrieval time and space complexity. This paper proposed deep multi-supervised hashing(DMSH) to learn highly discriminative and compact hash code for efficient image retrieval. We desinged a new CNN architecture to produce similarity-perserving hash code. A recognition signal was used to increase the inter-class distance and a verification signal was used to reduce the intra-class distance. We also applied regularization to reduce the loss between network output and binary hash code, and the binary hash value was evenly distributed in each dimension to make the network output closer to the discrete hash value. The experiments conducted on two datasets demonstrate that our proposed DMSH can fast encode any new-coming images and yield the state-of-the-art retrieval performance.
下载PDF全文