摘要
利用生成式对抗网络为零样本分类的问题提供新的思路.传统的生成式对抗网络是以不可见类的文本为原型生成新的不可见类产生相关样本的视觉信息.在传统生成式对抗网络中引入可见类文本信息,将可见类文本信息和不可见类文本信息对齐,再将对齐后不可见类文本信息生成不可见类视觉信息样本,以双向监督方式来解决零样本问题.在公开数据集模型上进行验证,数据集CUB、SUN、AWA的准确率分别提升了3.3%、1.1%、4.9%,结果表明双向监督的生成式对抗网络解决零样本分类的有效性.
Using generative adversarial networks,it provides new ideas for the problem of zero-sample classification.Traditional generative adversarial networks use the text of unseen classes as a prototype to generate new unseen classes to generate visual information about relevant samples.However,the model in this paper introduces visible text information into the traditional generative adversarial network,aligns the visible text information with the invisible text information,and then generates invisible visual information samples from the aligned invisible text information for bidirectional supervision.way to solve the above zero-sample problem.Validated on public datasets,the datasets CUB,SUN,and AWA have improved by 3.3%,1.1%,and 4.9%,respectively.The results show that the bidirectionally supervised generative adversarial network is effective in solving zero-sample classification.
作者
张伟
ZHANG Wei(Institute of Science And Technology,Changzhou Open University,Changzhou 213001,China)
出处
《南京工程学院学报（自然科学版）》
2022年第3期33-37,共5页
Journal of Nanjing Institute of Technology(Natural Science Edition)
