【摘要】 典型相关分析是多视图特征学习领域的研究热点,然而监督信息的缺失使其难以学习强鉴别力的相关特征,为此本文提出了一种新的鉴别相关特征学习方法,即标签散布相关分析(Label Scatter Correlation Analysis,LSCA)。该方法借助类标签信息,最大化了视图间类内相关性,并且最小化了视图间类间相关性和视图内类内散布,进而学习的相关特征在最大化相关性同时,尽可能的保留了类标签的鉴别力和散布结构。良好的实验结果已经显示该方法在图像识别中的有效性。更多还原

【Abstract】 Canonical correlation analysis(CCA) is a hot research in multi-view feature learning. However, due to the lack of supervised information, CCA is difficult to obtain correlation features with well discrimination power. To solve this issue, we propose a novel discriminant correlation feature learning method, i.e. label scatter correlation analysis(LSCA). By means of class label information, the method maximizes intra-class correlations between different views, and minimizes between-view inter-class correlations and within-view intra-class scatters. Thus correlation features learned by our method not only consider the maximum of between-view correlations but also further preserve the discrimination power of class labels and the scatter structures. Encouraging experimental results has showed the effectiveness of the method.更多还原