基于属性关系图正则化特征选择的零样本分类

Zero-shot classification based on attribute correlation graph regularized feature selection

为挖掘属性学习中属性与特征、属性与属性之间的关系,针对属性学习中存在的所有特征与属性被同等对待,底层特征与属性、属性与属性之间的先验知识被忽略的问题,提出一种基于属性关系图正则化特征选择的零样本分类方法.首先,根据训练样本和类别-属性矩阵计算属性之间的正负相关性,进而构建属性关系图;然后,基于属性关系图,对底层特征进行图正则化特征选择,并将选择后的特征用于直接属性预测(DAP)模型的训练;最后,通过直接属性分类器对测试样本进行零样本分类.AWA数据集上的实验结果表明,在40类训练10类测试的情况下,所提方法获得了0.692 6的属性预测平均AUC值及19.5%的零样本分类精度.

To explore the relationships of attribute-feature and attribute-attribute in attribute learning,and to solve the problems of all low-level features and attributes being treated equally and the prior knowledge about feature-attribute and attribute-attribute being ignored,a zeroshot classification method based on attribute correlation graph regularized feature selection was proposed.Firstly,based on training samples and the known class-attribute matrix,the positive or negative correlation between attributes could be computed and thus the attribute correlation graph could be built.Secondly,based on the attribute correlation graph,graph regularized feature selection was performed on the low-level features and the selected features were then used to train a direct attribute classifier（DAP）.Finally,zero-shot classification of testing samples was performed by the trained attribute classifier.Experimental results on the AWA dataset show that,under 40 training and 10 testing classes,the mean AUC value of attribute prediction and the zero-shot classification accuracy of the proposed method reach 0.692 6and 19.5%respectively.