选择性集成学习多判别器生成对抗网络

doi:10.3778/j.issn.1673-9418.2011010

计算机科学与探索 ›› 2022, Vol. 16 ›› Issue (6): 1429-1438.DOI: 10.3778/j.issn.1673-9418.2011010

• 图形图像 • 上一篇

选择性集成学习多判别器生成对抗网络

申瑞彩¹^,², 翟俊海¹^,²^,⁺(), 侯璎真¹^,²

1. 河北大学数学与信息科学学院,河北保定 071002
2. 河北大学河北省机器学习与计算智能重点实验室,河北保定 071002

收稿日期:2020-11-04 修回日期:2021-01-05 出版日期:2022-06-01 发布日期:2021-01-28
通讯作者: + E-mail: mczjh@126.com
作者简介:申瑞彩（1993—）,女,河北邯郸人,硕士研究生,主要研究方向为深度学习。
翟俊海（1964—）,男,河北易县人,博士,教授,主要研究方向为机器学习、深度学习、大数据处理。
侯璎真（1995—）,女,河北唐山人,硕士研究生,主要研究方向为深度学习。
基金资助:
河北省科技计划重点研发项目(19210310D);河北省自然科学基金(F2021201020)

Multi-discriminator Generative Adversarial Networks Based on Selective Ensemble Learning

SHEN Ruicai¹^,², ZHAI Junhai¹^,²^,⁺(), HOU Yingzhen¹^,²

1. College of Mathematics and Information Science, Hebei University, Baoding, Hebei 071002, China
2. Hebei Key Laboratory of Machine Learning and Computational Intelligence, Hebei University, Baoding, Hebei 071002, China

Received:2020-11-04 Revised:2021-01-05 Online:2022-06-01 Published:2021-01-28
About author:SHEN Ruicai, born in 1993, M.S. candidate. Her research interest is deep learning.
ZHAI Junhai, born in 1964, Ph.D., professor. His research interests include machine learning, deep learning and big data processing.
HOU Yingzhen, born in 1995, M.S. candidate. Her research interest is deep learning.
Supported by:
Key Research and Development Project of Hebei Provincial Science and Technology Plan(19210310D);Natural Science Foundation of Hebei Province(F2021201020)

摘要/Abstract

摘要：

生成对抗网络（GAN）在图像生成方面具有广泛应用,但基于无监督方式与有监督方式的网络生成样本仍有较大差距。为解决生成对抗网络在无监督环境中生成样本多样性差、质量较低以及模型训练时间过长等问题,提出了具有选择性集成学习思想的生成对抗网络模型。将生成对抗网络中的判别网络采用集成判别系统的形式,有效减少了由单判别器判别性能不佳导致判别误差的情况;同时考虑到若集成判别网络均采用统一网络设置,则在模型训练中基判别网络将趋近于一种表现形式,为鼓励判别网络判别结果多样且避免网络陷入雷同,设置拥有不同网络结构的判别网络,并在集成判别网络中引入具有动态调整基判别网络投票权重的多数投票策略,对集成判别网络的判别结果进行投票,有效地促进了模型的收敛且较大减少了实验误差。最后将提出的模型与同方向的模型在不同数据集上使用不同评价指标进行评价,实验结果表明提出的模型无论在生成样本多样性、生成样本质量还是在模型收敛速度上均明显优于几种竞争模型。

关键词: 生成对抗网络（GAN）, 集成判别系统, 选择性集成学习, 多数投票策略

Abstract:

Generative adversarial networks (GAN) are widely used in image generation. However, there is still a big gap between the samples generated by unsupervised and supervised networks. In order to solve the problems such as poor diversity, low quality and long training time of GAN in unsupervised environment, a new model with selective ensemble learning is proposed. Specifically, the discriminator in GAN is adopted in the form of integrated discrimination system, which can effectively reduce the discrimination error caused by the poor performance of single discriminator. Considering that if the integrated discriminant networks are set up in a unified network, each base discriminant network will tend to a form of expression in the model training. In order to encourage the diversity of discriminant network results and avoid the network falling into the same one, the discriminant networks with different network structures are set up. The majority voting strategy with dynamically adjusting the voting weight of the base discriminant network is introduced to vote the results of the integrated discriminant network. This has been shown to be effective in promoting model convergence and reducing experimental error significantly. Finally, the proposed model and the models in the same direction are evaluated with different evaluation indices under different datasets. Experimental results show that the proposed model is superior to several competitive models in terms of the diversity of generated samples, the quality of generated samples and the convergence speed of the model.

Key words: generative adversarial networks (GAN), integrated discrimination system, selective ensemble learning, majority voting strategy

中图分类号:

TP181

申瑞彩, 翟俊海, 侯璎真. 选择性集成学习多判别器生成对抗网络[J]. 计算机科学与探索, 2022, 16(6): 1429-1438.

SHEN Ruicai, ZHAI Junhai, HOU Yingzhen. Multi-discriminator Generative Adversarial Networks Based on Selective Ensemble Learning[J]. Journal of Frontiers of Computer Science and Technology, 2022, 16(6): 1429-1438.

图/表 21

图1 生成对抗网络模型

Fig.1 Model of generative adversarial networks

图2 选择性集成学习示意图

Fig.2 Diagram of selective ensemble learning

图3 选择性集成学习多判别器生成对抗网络模型

Fig.3 Selective ensemble learning for multi-discriminator generative adversarial network model

表1 针对MNIST数据集的各网络卷积层配置

Table 1 Convolution layers configuration of each discriminant network for MNIST dataset

网络	Con1	Con2	Con3	Con4
EGAN-3	（28,28,1）	（14,14,32/16/8）	（7,7,64/32/16）	（4,4,128/64/32）
EGAN-4	（28,28,1）	（14,14,64/32/16/8）	（7,7,128/64/32/16）	（4,4,256/128/64/32）
EGAN-5	（28,28,1）	（14,14,128/64/32/16/8）	（7,7,256/128/64/32/16）	（4,4,512/256/128/64/32）
EGAN-6	（28,28,1）	（14,14,256/128/64/32/16/8）	（7,7,512/256/128/64/32/16）	（4,4,1 024/512/256/128/64/32）

图4 采用不同集成策略生成的样本对比

Fig.4 Network generated samples by different integration strategies

表2 不同集成模型在两种集成策略下的IS得分

Table 2 IS score of different integration models under two integration strategies

集成策略	EGAN-3	EGAN-4	EGAN-5	EGAN-6
多数投票	1.819±0.314	1.916±0.263	2.010±0.206*	1.852±0.316
均值方式	1.736±0.247	1.788±0.308	1.809±0.207	1.503±0.401

表3 不同集成模型在两种集成策略下的FID得分

Table 3 FID score of different integration models under two integration strategies

集成策略	EGAN-3	EGAN-4	EGAN-5	EGAN-6
多数投票	3.872	3.360	3.027*	3.618
均值方式	4.304	4.588	4.562	4.621

表4 不同集成模型在两种集成策略下的KID得分

Table 4 KID score of different integration models under two integration strategies

集成策略	EGAN-3	EGAN-4	EGAN-5	EGAN-6
多数投票	31.65±1.52	30.42±1.49	29.98±1.46*	31.21±1.40
均值方式	34.51±1.72	34.33±1.32	34.21±1.49	35.02±1.54

图6 不同模型生成的手写体图像对比

Fig.6 Comparison of handwritten images generated by different models

图5 EGAN-n ( n = 3 , 4 , 5 , 6 )模型生成的手写体图像

Fig.5 Handwritten images generated by EGAN-n ( n = 3,4 , 5,6 )

表5 不同模型在三种评价指标下的得分情况

Table 5 Score of different models under three evaluation indices

模型	IS	FID	KID
GAN（集成）	1.201±0.040	4.351	50.96±1.73
DCGAN（集成）	1.572±0.310	3.780	37.64±1.49
GMAN	1.602±0.430	3.726	36.04±1.40
EGAN-5	2.010±0.206*	3.027*	29.98±1.46*

表6 针对CelebA数据集的各网络卷积层配置

Table 6 Convolution layers configuration of each discriminant network for CelebA dataset

层数	Con2	Con3	Con4	Con5	Con6
3	（80,64,16/8/4）	（40,32,32/16/8）	（20,16,64/32/16）	（10,8,128/64/32）	(5,4,256/128/64）
4	（80,64,32/16/8/4）	（40,32,64/32/16/8）	（20,16,128/64/32/16）	（10,8,256/128/64/32）	(5,4,512/256/128/64）
5	（80,64,64/32/16/8/4）	（40,32,128/64/32/16/8）	（20,16,256/128/64/32/16）	（10,8,512/256/128/64/32）	(5,4,1 024/512/256/128/64）
6	（80,64,128/64/32/16/8/4）	（40,32,256/128/64/32/16/8）	（20,16,512/256/128/64/32/16）	（10,8,1 024/512/256/128/64/32）	(5,4,512/256/256/128/128/64）

图7 采用不同集成策略生成的样本对比

Fig.7 Network generated samples by different integration strategies

表7 不同集成模型在两种集成策略下的IS得分

Table 7 IS score of different integration models under two integration strategies

集成策略	EGAN-3	EGAN-4	EGAN-5	EGAN-6
多数投票	2.716±0.630	3.242±0.512	3.413±0.426*	2.967±0.583
均值方式	2.391±0.232	2.502±0.374	2.401±0.246	2.304±0.913

表8 不同集成模型在两种集成策略下的FID得分

Table 8 FID score of different integration models under two integration strategies

集成策略	EGAN-3	EGAN-4	EGAN-5	EGAN-6
多数投票	1.381	1.196	1.094*	1.265
均值方式	1.649	1.588	1.701	1.647

表9 不同集成模型在两种集成策略下的KID得分

Table 9 KID score of different integration models under two integration strategies

集成策略	EGAN-3	EGAN-4	EGAN-5	EGAN-6
多数投票	0.402±0.496	0.204±0.451	0.196±0.216*	0.365±0.401
均值方式	0.594±0.824	0.571±0.763	0.497±0.801	0.734±0.671

图8 EGAN-n （ n = 3 , 4 , 5 , 6 ）模型生成的人脸图像

Fig.8 Face images generated by EGAN-n ( n = 3,4 , 5,6 ) model

图9 不同模型生成的人脸图像对比

Fig.9 Comparison of face images generated by different models

表10 不同模型在三种评价指标下的得分情况

Table 10 Score of different models under three evaluation indices

模型	IS	FID	KID
GAN（集成）	1.561±0.412	2.406	1.473±0.504
DCGAN（集成）	2.013±0.892	1.998	0.996±0.291
GMAN	2.541±0.340	1.942	0.983±0.274
EGAN-5	3.413±0.426*	1.094*	0.196±0.216*

图10 不同模型训练一个epoch所需的时间

Fig.10 Time required for different models to be trained in an epoch

图11 生成的部分样例（左）与数据集的部分样例（右）

Fig.11 Some generated samples (left) and some samples in dataset (right)

参考文献 19

[1]	GOODFELLOW I J, POUGET-ABADIE J, MIRZA M, et al. Generative adversarial nets[C]// Proceedings of the 27th International Conference on Neural Information Proces-sing Systems, Montréal, Dec 8-13, 2014. Cambridge: MIT Press, 2014: 2672-2680.
[2]	DURUGKER I, GEMP I, MAHADEVAN S. Generative multi-adversarial networks[EB/OL]. [2018-07-04]. https://open-review.net/forum?id=Byk-VI9eg .
[3]	兰远东, 曾树洪. 动态加权投票的多分类器聚合[J]. 现代计算机(专业版), 2014(2): 8-11.
	LAN Y D, ZENG S H. Multiple classifier fusion with dyna-mic weighted voting[J]. Modern Computer, 2014(2): 8-11.
[4]	SCHUTZ B, ILYAS S, LE K, et al. Nanoparticle arrays having directed hybrid topology via covalent self-assembly of iron oxide and silica nanoparticles[J]. ACS Applied Nano Materials, 2020, 3(6): 5936-5943. DOI URL
[5]	CHAKRABORTY S, ROY M. A multi-level weighted trans-formation based neuro-fuzzy domain adaptation technique using stacked auto-encoder for land-cover classification[J]. International Journal of Remote Sensing, 2020, 41(17): 6831-6857. DOI URL
[6]	ZORAN J, NICOLA C, PRATS D B, et al. A highly parame-terizable framework for conditional restricted Boltzmann machine based workloads accelerated with FPGAs and OpenCL[J]. Future Generation Computer Systems, 2020, 104: 201-211. DOI URL
[7]	ZHENG J, SU Y X, ZHANG D H, et al. Velocity forecasts using a combined deep learning model in hybrid electric vehicles with V2V and V2I communication[J]. Science China Technological Sciences, 2020, 63(1): 55-64. DOI URL
[8]	HAN Z, TAO X, HONGSHENG L, et al. StackGAN++: rea-listic image synthesis with stacked generative adversarial networks[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019, 41(8): 1947-1964. DOI URL
[9]	ZHU J Y, PARK T, ISOLA P, et al. Unpaired image-to-image translation using cycle-consistent adversarial net-works[C]// Proceedings of the 2017 IEEE International Conference on Computer Vision, Venice, Oct 22-29, 2017. Washington: IEEE Computer Society, 2017: 2242-2251.
[10]	HOFFMAN J, TZENG E, PARK T, et al. Cycle-consistent adversarial domain adaptation[C]// Proceedings of the 2017 IEEE Conference on Computer Vision and Pattern Recog-nition, Hawaii, Jul 21-26, 2017. Piscataway: IEEE, 2017: 2962-2971.
[11]	HUANG H, LIU M Y, BELONGIE S J, et al. Multimodal unsupervised image-to-image translation[C]// LNCS 11207: Proceedings of the 15th European Conference on Computer Vision, Munich, Sep 8-14, 2018. Cham: Springer, 2018: 179-196.
[12]	RADFOD A, METZ L, CHINTALA S. Unsupervised rep-resentation learning with deep convolutional generative ad-versarial networks[J]. Computer Science, 2015, 47(8): 169-183.
[13]	CHEN X, DUAN Y, HOUTHOOFT R, et al. Info-GAN: in-terpreter representation learning by information maximi-zing generative adversarial nets[C]// Proceedings of the 30th International Conference on Neural Information Processing Systems, Barcelona, Dec 5-10, 2016. Red Hook: Curran As-sociates, 2016: 2172-2180.
[14]	方敏. 集成学习的多分类器动态融合方法研究[J]. 系统工程与电子技术, 2006, 28(11): 1759-1761.
	FANG M. Study of integration method for multiple classifiers on ensemble learning[J]. Systems Engineering and Elec-tronics, 2006, 28(11): 1759-1761.
[15]	LIU Y M, TUZEL O. Coupled generative adversarial net-works[C]// Proceedings of the 30th International Confe-rence on Neural Information Processing Systems, Barce-lona, Dec 5-10, 2016. Red Hook: Curran Associates, 2016: 469-477.
[16]	周志华. 机器学习[M]. 北京: 清华大学出版社, 2016: 171-196.
	ZHOU Z H. Machine learning[M]. Beijing: Tsinghua Uni-versity Press, 2016: 171-196.
[17]	郑哲, 胡庆浩, 刘青山, 等. 量化权值激活的生成对抗网络[J]. 计算机科学, 2020, 47(5): 144-148.
	ZHENG Z, HU Q H, LIU Q S, et al. Quantizing weights and activations in generative adversarial networks[J]. Com-puter Science, 2020, 47(5): 144-148.
[18]	王坤峰, 苟超, 段艳杰, 等. 生成式对抗网络GAN的研究进展与展望[J]. 自动化学报, 2017, 43(3): 321-332.
	WANG K F, GOU C, DUAN Y J, et al. Generative adver-sarial networks: the state of the art and beyond[J]. Acta Automatica Sinica, 2017, 43(3): 321-332.
[19]	张春霞, 张讲社. 选择性集成学习算法综述[J]. 计算机学报, 2011, 34(8): 1399-1410.
	ZHANG C X, ZHANG J S. A survey of selective ensemble learning algorithms[J]. Chinese Journal of Computers, 2011, 34(8): 1399-1410. DOI URL

选择性集成学习多判别器生成对抗网络

Multi-discriminator Generative Adversarial Networks Based on Selective Ensemble Learning

RichHTML

PDF

可视化

摘要/Abstract

引用本文

使用本文

图/表 21

参考文献 19

相关文章 13

编辑推荐

Metrics

[1]	夏鸿斌, 肖奕飞, 刘渊. 融合自注意力机制的长文本生成对抗网络模型[J]. 计算机科学与探索, 2022, 16(7): 1603-1610.
[2]	林佳伟, 王士同. 用于无监督域适应的深度对抗重构分类网络[J]. 计算机科学与探索, 2022, 16(5): 1107-1116.
[3]	姜艺, 胥加洁, 柳絮, 朱俊武. 边缘指导图像修复算法研究[J]. 计算机科学与探索, 2022, 16(3): 669-682.
[4]	王曙燕, 金航, 孙家泽. GAN图像对抗样本生成方法[J]. 计算机科学与探索, 2021, 15(4): 702-711.
[5]	费建伟，夏志华，余佩鹏，戴昀书. 人脸合成技术综述[J]. 计算机科学与探索, 2021, 15(11): 2025-2047.
[6]	王富平, 李文楼, 刘颖, 卢津, 公衍超. 结合边缘信息和门卷积的人脸修复算法[J]. 计算机科学与探索, 2021, 15(1): 150-162.
[7]	彭晏飞，高艺，杜婷婷，桑雨，訾玲玲. 生成对抗网络的单图像超分辨率重建方法[J]. 计算机科学与探索, 2020, 14(9): 1612-1620.
[8]	李广丽，滑瑾，袁天，朱涛，邬任重，姬东鸿，张红斌. 基于用户偏好挖掘生成对抗网络的推荐系统[J]. 计算机科学与探索, 2020, 14(5): 803-814.
[9]	武随烁，杨金福，单义，许兵兵. 使用孪生注意力机制的生成对抗网络的研究[J]. 计算机科学与探索, 2020, 14(5): 833-840.
[10]	李俊杰，王茜. 感知相似的图像分类对抗样本生成模型[J]. 计算机科学与探索, 2020, 14(11): 1930-1942.
[11]	梁俊杰，韦舰晶，蒋正锋. 生成对抗网络GAN综述[J]. 计算机科学与探索, 2020, 14(1): 1-17.
[12]	商显震，韩萌，孙毓忠，孙宇宁，陈旭，胡满满，梅御东. 融合生成对抗网络和朴素贝叶斯皮肤病诊断方法[J]. 计算机科学与探索, 2019, 13(6): 1005-1015.
[13]	许欧阳，李光辉. 萤火虫优化和随机森林的WSN异常数据检测[J]. 计算机科学与探索, 2018, 12(10): 1633-1644.