基于深度学习特征的图像检索方法  

Image retrieval method via deep learning features

作　　者：任夏荔 陈光喜 曹建收 蔡天任 

机构地区：[1]桂林电子科技大学计算机与信息安全学院,广西桂林541004

出　　处：《计算机工程与设计》2018年第2期503-510,共8页Computer Engineering and Design

基　　金：广东省数学教育软件工程技术研究中心开放基金项目(LD16124X);广西自然科学基金项目(2013GXNSFC019330);广西图像处理重点实验室基金项目(LD15042X);广西学位与研究教育改革基金项目(JGY2014060)

摘　　要：针对图像高维特征占用巨大存储空间的问题,考虑到其中一些特征存在高度相关性,提出一种基于深度学习特征的图像检索方法。利用卷积神经网络,提取基于深度学习的图像特征;通过分析特征本身存在的相关性,采用主成分分析算法,对特征降维的同时尽量降低信息的损失;在此基础上对特征进行哈希编码,利用编码做快速图像检索。Caltech101和Caltech256数据集上的实验结果表明,相比直接使用原始深度特征进行编码检索,所提方法的检索性能更好。As for the problem that the high dimensional image’s feature uses large storage space,and considering that some of these features have a high degree of correlation,a principal components analysis via deep learning features based on image retrieval method was proposed.Deep learning features were extracted based on convolutional neural network.By analyzing the correlation of features and using principal component analysis algorithm,dimensions of the features were reduced while information loss was minimized.Binary hash codes were used for fast image retrieval.Experimental results show that the retrieval perfor-mance is better than the property of the method that using all of the untreated characteristics to encode and retrieval on the Caltech101 and Caltech256 datasets.

关 键 词：图像检索 卷积神经网络 深度学习 主成分分析 哈希编码 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]