%0 Journal Article
%T 利用深度玻尔兹曼机与典型相关分析的自动图像标注算法<br>An Automatic Image Annotation Algorithm Using Deep Boltzmann Machine and Canonical Correlation Analysis
%A 刘凯
%A 张立民
%A 孙永威
%A 林雪原
%J 西安交通大学学报
%D 2015
%R 10.7652/xjtuxb201506006
%X 提出一种基于深度玻尔兹曼机与典型相关分析的自动图像标注算法(DBM？？CCA)。该算法利用深度玻尔兹曼机实现图像与文本的低层次特征向稀疏高层次抽象概念的转变,并通过典型相关分析建立子空间映射关系以实现标注词汇的生成。首先在深度玻尔兹曼机提取图像与文本高层特征过程中,选用伯努利分布和高斯分布分别拟合标注词汇和图像特征,然后在图像与标注词汇高层特征形成的典型变量空间内计算待标注图像与训练集图像的马氏距离并据此加权计算得到高层标注词汇特征,最后由平均场估计生成图像标注词汇。实验结果表明,所提算法对图像的标注准确率改善较好,与经典的基于监督的多类标签方法和多重伯努利相关模型相比,在Corel5K实验中平均查准率和查全查准均率分别提高了10％和5％。<br>An automatic image annotation algorithm is proposed based on deep Boltzmann machine and canonical correlation analysis, named DBM？？CCA. The algorithm utilizes DBM to transform low？？level features of images and labels to sparse high？？level abstract concepts, and builds subspace mapping relations by CCA in order to generate labels. The multiple Bernoulli distribution is used to fit labels and the Gaussian distribution is used to fit image features in the process of using DBM to extract high？？level features of images and labels. CCA is used to establish relevant connection among image features and labeling words which form canonical variable subspace. High？？level text features are calculated based on the Mahalanobis distance between images in canonical variable subspace, and image annotation words are generated by mean？？field inference. Experimental results show that the proposed automatic image annotation method significantly outperforms both the traditional MBRM and the SML, and the precision ratio and recall？？precision mean ratio are increased by 10％ and 5％, respectively, in experiments with Corel5K image dataset
%K 自动图像标注
%K 深度学习
%K 深度玻尔兹曼机
%K 典型相关分析&lt
%K br&gt
%K automatic image annotation
%K deep learning
%K deep Boltzmann machine
%K canonical correlation analysis
%U http://zkxb.xjtu.edu.cn/oa/DArticle.aspx?type=view&id=201506006