基于共享特征相对属性的零样本图像分类

Shared Features Based Relative Attributes for Zero-shot Image Classification

在利用相对属性学习实现零样本图像分类中,现有的方法并没有考虑属性与类别之间的关系,为此该文提出一种基于共享特征相对属性的零样本图像分类方法。该方法采用多任务学习的思想来共同学习类别分类器和属性分类器,获得一个低维的共享特征子空间,挖掘属性与类别之间的关系。同时,利用共享特征来学习属性排序函数,得到基于共享特征的相对属性模型,解决了相对属性学习过程中丢失属性与类别关系的问题。另外,将基于共享特征的相对属性模型用于零样本图像分类中,有效提高了零样本图像分类的识别率。实验数据集上的结果表明,该方法具有较高的相对属性学习性能和零样本图像分类精度。

Most algorithms of the zero-shot image classification with relative attributes do not consider the relationship between attributes and classes, therefore a new relative attributes method based on shared features is proposed for zero-shot image classification. In analogy to the multi-task learning, the object classifier and attribute classifier are simultaneously learned in this method, from which a shared sub-space of lower dimensional features is obtained to mine the relationship between attributes and classes. Inspired by the success of shared features, a novel relative attributes model based on shared features is proposed to promote the performance of the relationship between attributes and classes, in which the ranking function per attribute is learned by using shared features. In addition, the novel relative attributes model based on shared features is applied to zero-shot image classification, which yields high accuracy due to the shared features included. Experimental results demonstrate that the proposed method can achieve high relative attributes learning efficiency and zero-shot image classification accuracy.