基于粒子群算法寻最优属性关联下的零样本语义自编码器

针对零样本图像分类构建共享属性层时造成的信息缺失问题,该文提出一种嵌入属性关联性的补偿方法.通过语义自编码器构建特征到属性的映射,然后以最大后验概率估计在类高斯模型构建的基础上实现零样本图像分类.为弥补SAE对属性关系学习的不足,引入加性因子与乘性因子对属性相关性进行嵌入,并利用粒子群算法搜寻最优的因子参数,实现属性相关性信息的补偿.实验结果表明采取相同映射方法的情况下,基于属性相关性嵌入的零样本图像分类在Pubfig数据集和OSR数据集上的分类效果较之其他方法得到了显著提升.

Zero-shot Learning by Semantic Autoencoder Based on Particle Swarm Optimization Algorithm for Attribute Correlation

To deal with the problem of missing information caused by zero-shot image classification during building a shared attribute layer, a compensation method is proposed to embed the attribute correlation. The proposed zero-shot classification utilizes Semantic AautoEncoder (SAE) to realize the feature-to-attribute mapping, and the invisible images are classified using maximum posterior probability estimation based on the class Gaussian distribution model. In order to make up for the lack of attribute relationships in SAE learning, the additive and multiplicative factors are introduced to embed the attribute correlation. The particle swarm algorithm is used to search for the optimal factor parameters to achieve the compensation of attribute correlation information. Experimental results show that when the same mapping method is adopted, the classification performance of zero-shot image classification based on attribute correlation on Pubfig and OSR data sets is significantly improved compared with other methods.