零样本图像分类方法综述

A Survey of Zero-Shot Image Classification Methods

图像分类作为图像处理的基础,一直是热门的研究话题。传统的图像分类方法需要采用大量专家标记的数据集,成本较大且无法识别未训练的图像类别;零样本图像分类作为迁移学习的一个分支,可以将已有的知识属性迁移到未知类中,从而完成对未知类图像的分类识别,因此降低了训练数据的标记成本,且能在样本稀缺的情况下实现对新事物的分类识别。首先,简单介绍了零样本学习的概念;其次,着重介绍了基于属性预测、基于特征嵌入和基于生成模型的零样本图像分类方法;然后,简要介绍了零样本图像分类的数据集以及评估方法,并对经典模型的实验结果做了比较分析;最后,提出了零样本图像分类方法普遍存在的问题以及相应的解决思路。

Image classification,as the foundation of image processing,has always been a hot research topic.Traditional image classification methods use supervised learning,which requires a large number of expert labeled datasets,this is costly and cannot recognize untrained image categories.As a branch of transfer learning,zero-shot image classification can transfer existing knowledge attributes to unknown classes,thus completing the classification and recognition of unknown images.Therefore,it reduces the marking cost of training data,and can realize the classification and recognition of new images with few or no samples.First,the zero-shot learning is briefly introduced.Second,the zero-shot image classification method based on attribute prediction.The zero sample image classification method based on feature embedding.And the zero sample image classification method based on generative model are emphatically introduced.Third,the dataset and evaluation method of zero-shot image classification are briefly introduced,and the experimental results of classical models are compared and analyzed.Finally,the common problems and corresponding solutions of zero sample image classification methods were discussed.