%0 Journal Article %T 基于优选典型相关分量的跨媒体检索模型<br>Cross-media retrieval model based on choosing key canonical correlated vectors %A 李广丽 %A 刘斌 %A 朱涛 %A 殷依 %A 张红斌< %A br> %A Guangli LI %A Bin LIU %A Tao ZHU %A Yi YIN %A Hongbin ZHANG %J 山东大学学报（工学版） %D 2018 %R 10.6040/j.issn.1672-3961.0.2017.552 %X 在跨媒体检索中,准确利用异构媒体间的语义相关性是制约检索性能优劣的关键因素之一。提出改进的核典型相关分析(modified kernel canonical correlation analysis, MKCCA)模型,以改善跨媒体检索性能:抽取图像的尺度不变特征变换(scale invariant feature transform, SIFT)与描述灰度纹理的空间包络特征(GIST),抽取文本的词频(term frequency, TF)特征;精选映射核,把图像、文本特征映射到高维可分空间中,生成核矩阵;基于典型相关分析(canonical correlation analysis, CCA)方法挖掘图像、文本核矩阵间的非线性语义相关性;设置语义相关度阈值,降低语义噪声干扰并优选核心典型相关分量,更准确、鲁棒地刻画图像与文本间的语义关联。试验表明:SIFT-TF特征组合整体表现最好,而MKCCA模型与高斯核(gauss kernel)配合可获取最优跨媒体检索性能,其图像检索文本与文本检索图像的平均精度均值(mean average precision, MAP)较次优指标分别提升3.06%和1.18%。<br>It is one of the most important factors which affect final retrieval performance effectively by acquiring the core semantic correlations between heterogeneous media in cross-media retrieval. To improve retrieval performance, a modified kernel canonical correlation analysis (MKCCA) model was presented: image features like SIFT (scale invariant feature transform) and GIST were extracted respectively to better characterize the key visual content of images. Meanwhile TF (term frequency) feature was extracted to depict the key characteristics of texts. Then the extracted features were mapped into a high-dimensional space by mapping kernels. As the results, two kernel matrixes were acquired to describe the mapped features. Based on the kernel matrixes, the non-linear semantic correlations between images and texts were fully mined by canonical correlation analysis (CCA) model. More importantly, with the help of a semantic correlation threshold, those core canonical correlation vectors were chosen to suppress semantic noises and depict the key semantic correlations between images and texts more robustly. Experimental results showed that the best overall retrieval performance was obtained by using the feature combination SIFT-TF. Moreover the highest retrieval performance was obtained by MKCCA model combined with gauss kernel. Compared to the best competitor, the MAP value of the "images retrieve texts (I_R_T)" task was improved about 3.06% while the MAP value of the "texts retrieve image (T_R_I)" task was improved about 1.18%. %K 典型相关分量 %K 跨媒体检索 %K 核典型相关分析 %K 语义相关度阈值 %K 高斯核 %K < %K br> %K canonical correlated vectors %K cross-media retrieval %K kernel canonical correlation analysis %K semantic correlation threshold %K gauss kernel %U http://gxbwk.njournal.sdu.edu.cn/CN/10.6040/j.issn.1672-3961.0.2017.552