基于条件梯度Wasserstein生成对抗网络的图像识别  

Image Recognition With Conditional Wasserstein Generative Adversarial Networks with Gradient Penalty

作　　者：何子庆 聂红玉 刘月[1] 尹洋 He Ziqing;Nie Hongyu;liu Yue;Yin Yang(College of Information Science and Technology,Southwest Jiaotong University,Chengdu 610097,China;School of Big Data,Chongqing Vocational College of Transportation,Chongqing 402247,China)

机构地区：[1]西南交通大学信息科学与技术学院,成都610097 [2]重庆交通职业学院大数据学院,重庆402247

出　　处：《计算机测量与控制》2019年第6期157-162,共6页Computer Measurement &Control

基　　金：国家自然科学基金资助项目(61461048);重庆市教委科学技术研究项目(KJQN201805702);四川省科技创新苗子工程资助项目(2018102)

摘　　要：生成式对抗网络GAN功能强大,但是具有收敛速度慢、训练不稳定、生成样本多样性不足等缺点;该文结合条件深度卷积对抗网络CDCGAN和带有梯度惩罚的Wasserstein生成对抗网络WGAN-GP的优点,提出了一个混合模型-条件梯度Wasserstein生成对抗网络CDCWGAN-GP,用带有梯度惩罚的Wasserstein距离训练对抗网络保证了训练稳定性且收敛速度更快,同时加入条件c来指导数据生成;另外为了增强判别器提取特征的能力,该文设计了全局判别器和局部判别器一起打分,最后提取判别器进行图像识别;实验结果证明,该方法有效的提高了图像识别的准确率。Generated adversarial net GAN is powerful,but it has some disadvantages such as slow convergence,unstable training,and insufficient sample diversity.This paper presents a conditional gradient Wasserstein generation confrontation network model CDCWGAN-GP by Combining the advantage of conditional deep convolution adversarial net CDCGAN and Wasserstein generated adversarial net with gradient penalty WGAN-GP.Using the Wasserstein distance training against the network with gradient penalty guarantees training stability and faster convergence,while adding condition c to guide data generation.In addition,in order to enhance the ability of the discriminator to extract features,the paper designs a global discriminator and a local discriminator to score together,and finally extracts the discriminator for image recognition.The result of simulation experiments show that this method effectively improves the accuracy of image recognition.

关 键 词：生成式对抗网络 条件模型 Wesserstein距离 梯度惩罚 全局和局部一致性 图像识别 

分 类 号：TP311.53[自动化与计算机技术—计算机软件与理论]