融合VAE和StackGAN的零样本图像分类方法  

Zero-shot image classification method combining VAE and StackGAN

作　　者：张冀[1] 曹艺 王亚茹 赵文清[1] 翟永杰[2] ZHANG Ji;CAO Yi;WANG Yaru;ZHAO Wenqing;ZHAI Yongjie(Department of Computer,North China Electric Power University,Baoding 071003,China;Department of Automation,North China Electric Power University,Baoding 071003,China)

机构地区：[1]华北电力大学计算机系,河北保定071003 [2]华北电力大学自动化系,河北保定071003

出　　处：《智能系统学报》2022年第3期593-601,共9页CAAI Transactions on Intelligent Systems

基　　金：国家自然科学基金面上项目(61773160);河北省自然科学基金青年科学基金项目(F2021502008);中央高校基本科研业务费专项资金面上项目(2021MS081).

摘　　要：零样本分类算法旨在解决样本极少甚至缺失类别情况下的分类问题。随着深度学习的发展,生成模型在零样本分类中的应用取得了一定的突破,通过生成缺失类别的图像,将零样本图像分类转化为传统的基于监督学习的图像分类问题,但生成图像的质量不稳定,如细节缺失、颜色失真等,影响图像分类准确性。为此,提出一种融合变分自编码(variational auto-encoder,VAE)和分阶段生成对抗网络(stack generative adversarial networks,StackGAN)的零样本图像分类方法,基于VAE/GAN模型引入StackGAN,用于生成缺失类别的数据,同时使用深度学习方法训练并获取各类别的句向量作为辅助信息,构建新的生成模型stc-CLS-VAEStackGAN,提高生成图像的质量,进而提高零样本图像分类准确性。在公用数据集上进行对比实验,实验结果验证了本文方法的有效性与优越性。The zero-shot classification algorithm is designed to solve the classification problem in case of a few samples or even missing categories.With the development of deep learning,the application of the generation model in zero-shot classification has made a breakthrough.By generating images of missing categories,the zero-shot image classification is transformed into a traditional image classification problem based on supervised learning.However,the generated samples are unstable in quality,including missing details and color distortion,thus affecting the accuracy of image classification.To this end,the zero-shot image classification method combining variational auto-encoding(VAE)and stack generative adversarial networks(StackGAN)is proposed.Based on the VAE/GAN model,StackGAN is introduced to generate the data of missing categories.Meanwhile,the deep learning method is used to train and obtain the sentence vectors of each category as auxiliary information and build a new generation model stc–CLS–VAEStackGAN to improve the quality of generated images and subsequently improve the classification accuracy of the zero-shot images.A comparative experiment was conducted on the public dataset,and the experimental results verified the effectiveness and superiority of the method proposed herein.

关 键 词：深度学习 零样本学习 图像分类 变分自编码器 生成对抗网络 分阶段网络 句向量 辅助信息 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]