基于反向投影的zero-shot learning目标分类算法研究

Zero-shot learning object recognition based on inverse projection

Zero-shot learning(ZSL)是针对没有训练样本的类别进行分类的问题。传统回归方法的核心是将视觉特征投影到语义空间,没有充分利用视觉特征自身包含的样本信息,同时训练计算量大。提出基于反向投影的ZSL目标分类方法,将类别原型投影到视觉空间,利用视觉特征的语义性学习出映射函数,参数优化过程仅通过解析解就可以获得。在两个基准数据集的实验结果表明,提出的反向投影方法分类结果较传统回归方法和其他现有方法有大幅提升,并且训练时间大大减少,可以更好地推广到未知类别的分类问题上。

Zero-shot learning（ZSL） aims to reeognise new objects without having training samples of them. The kernel of the traditional regression method was to project the visual features into the semantic space,without taking full advantage of the sample information contained in the visual features, meanwhile the training computation was large. This paper proposed to project prototype to visual feature space,which was referred to as inverse projection. It had a very. efficient elosed-furm solution. Extensive experiments on two benchmark datasets show thai the proposed ZSL method significantly outperforms the state-of-the-arts.