生成式对抗网络及其在图像生成中的研究进展

doi:10.3778/j.issn.1673-9418.2103075

摘要/Abstract

摘要：

生成式对抗网络（GAN）现已成为深度学习领域热门的研究方向，其独特的对抗性思想来源于博弈论中的二人零和博弈，如何解决GAN训练不稳定、生成样本质量差、评价体系不够健全、可解释性差等问题是目前GAN研究的重点和难点。调研了生成式对抗网络的研究背景和发展趋势。首先阐述了生成式对抗网络的基本思想和算法实现，分析了GAN的优势与不足，然后对已有改进方法进行了较为系统的分类，从基于结构改变和基于损失函数变体的两种类型分别梳理了一些典型的GAN的优化方法和衍生模型；比较了GAN与其他生成模型的异同，介绍了各自的优势与不足；对比了GAN及其衍生模型的性能，总结了它们的运作机制、优点、局限性以及适用场景，介绍了生成式对抗网络在图像生成领域中的应用；最后列举了生成式对抗网络的主流评价指标，分析了GAN研究中仍面临的主要问题并给出对应的解决思路，并将列举出的主流解决手段在解决效果及可应用性方面进行了对比分析，展望了未来的研究方向。

关键词: 生成式对抗网络（GAN）, 机器学习, 深度学习, 图像处理, 无监督学习, 图像生成

Abstract:

Generative adversarial networks (GAN) has become a popular research direction in the field of deep learning. The unique adversarial idea of GAN comes from the two-player zero-sum game in game theory. How to solve the problems of unstable GAN training, poor quality of generated samples, inadequate evaluation system, poor interpretability and other issues is critical and difficult in current GAN research. This paper investigates the research background and development trend of GAN. First, the basic idea and algorithm implementation of GAN are described, the advantages and disadvantages of GAN are analyzed, a more systematic classification of existing improved methods is made, and some typical optimization methods and derivative models of GAN are sorted out from two categories based on structure change and loss function variants respectively. Next, the similarities and differences between GAN and other generative models are compared, and their respective advantages and disadvantages are introduced. Then, this paper compares the performance of GAN and its derivative models, and summarizes their operation mechanism, advantages, limitations and applicable scenarios. The applications of GAN in the field of image generation are introduced. Finally, the mainstream evaluation indicators of GAN are listed, the main problems still faced in GAN research are analyzed and corresponding solutions are given. The listed mainstream solutions are compared and analyzed in terms of solution effects and applicability, and the future research direction is prospected.

Key words: generative adversarial networks (GAN), machine learning, deep learning, image processing, unsupervised learning, image generation

马永杰, 徐小冬, 张茹, 谢艺蓉, 陈宏. 生成式对抗网络及其在图像生成中的研究进展[J]. 计算机科学与探索, 2021, 15(10): 1795-1811.

MA Yongjie, XU Xiaodong, ZHANG Ru, XIE Yirong, CHEN Hong. Generative Adversarial Network and Its Research Progress in Image Generation[J]. Journal of Frontiers of Computer Science and Technology, 2021, 15(10): 1795-1811.

参考文献

[1] SHAO L, WU D, LI X. Learning deep and wide: a spectral method for learning deep networks[J]. IEEE Transactions on Neural Networks and Learning Systems, 2014, 25(12): 2303-2308.
[2] HINTON G E, SALAKHUTDINOV R R. Reducing the di-mensionality of data with neural networks[J]. Science, 2006, 313(5786): 504-507.
[3] GOODFELLOW I J, POUGET-ABADIE J, MIRZA M, et al. Generative adversarial networks[J]. arXiv:1406.2661, 2014.
[4] DOU K Z, CAI L Z M, NAN T B, et al. Tibetan speech syn-thesis based on neural network[J]. Journal of Chinese Infor-mation Processing, 2019, 33(2): 75-80.
都格草, 才让卓玛, 南措吉, 等. 基于神经网络的藏语语音合成[J]. 中文信息学报, 2019, 33(2): 75-80.
[5] XU D, WEI C, PENG P, et al. GE-GAN: a novel deep lear-ning framework for road traffic state estimation[J]. Transpor-tation Research Part C: Emerging Technologies, 2020, 117: 102635.
[6] WU Y, YANG F, XU Y, et al. Privacy-protective-GAN for privacy preserving face deidentification[J]. Journal of Com-puter Science and Technology, 2019, 34(1): 47-60.
[7] HU M F, LIU J W, ZUO X. Survey on deep generative model[J/OL]. Acta Automatica Sinica[2020-10-17]. https://doi.org/10.16383/j.aas.c190866.
胡铭菲, 刘建伟, 左信. 深度生成模型综述[J/OL]. 自动化学报[2020-10-17]. https://doi.org/10.16383/j.aas.c190866.
[8] WANG Z W, SHE Q, WARD T E. Generative adversarial networks in computer vision: a survey and taxonomy[J]. arXiv:1906.01529, 2019.
[9] MIRZA M, OSINDERO S. Conditional generative adversarial nets[J]. arXiv:1411.1784, 2014.
[10] ODENA A, OLAH C, SHLENS J. Conditional image syn-thesis with auxiliary classifier GANs[J]. arXiv:1610.09585, 2016．
[11] CHEN X, DUAN Y, HOUTHOOFT R, et al. InfoGAN: inter-pretable representation learning by information maximizing generative adversarial nets[J]. arXiv:1606.03657, 2016.
[12] PERARNAU G, VAN DE WEIJER J, RADUCANU B, et al. Invertible conditional GANs for image editing[J]. arXiv:1611.06355, 2016.
[13] DASH A, GAMBOA J C B, AHMED S, et al. TAC-GAN: text conditioned auxiliary classifier generative adversarial network[J]. arXiv:1703.06412, 2017.
[14] ODENA A. Semi-supervised learning with generative adver-sarial networks[J]. arXiv:1606.01583, 2016.
[15] SPRINGENBERG J T. Unsupervised and semi-supervised learning with categorical generative adversarial networks[J]. arXiv:1511.06390, 2015.
[16] RADFORD A, METZ L, CHINTALA S. Unsupervised repre-sentation learning with deep convolutional generative adver-sarial networks[J]. arXiv:1511.06434, 2015.
[17] XU B, WANG N Y, CHEN T Q, et al. Empirical evaluation of rectified activations in convolutional network[J]. arXiv:1505.00853, 2015.
[18] YU C C, KANG M, CHEN Y B, et al. Endangered Tujia language speech enhancement research based on improved DCGAN[C]//LNCS 11856: Proceedings of the 18th China National Conference on Chinese Computational Linguistics, Kunming, Oct 18-20, 2019. Cham: Springer, 2019: 394-404.
[19] ZHAO J B, MATHIEU M, LECUN Y. Energy-based genera-tive adversarial network[J]. arXiv:1609.03126, 2016.
[20] DENTON E L, CHINTALA S, SZLAM A, et al. Deep genera-tive image models using a Laplacian pyramid of adversarial networks[C]//Proceedings of the Annual Conference on Neural Information Processing Systems 2015, Montreal, Dec 7-12, 2015. Red Hook: Curran Associates, 2015: 1486-1494.
[21] OLKKONEN H, PESOLA P. Gaussian pyramid wavelet transform for multiresolution analysis of images[J]. Graphical Models & Image Processing, 1996, 58(4): 394-398.
[22] BURT P J，ADELSON E H. The Laplacian pyramid as a compact image code[J]. IEEE Transactions on Communica-tions, 1983, 31(4): 532-540.
[23] ZHANG H, XU T, LI H S, et al. StackGAN: text to photo-realistic image synthesis with stacked generative adversarial networks[C]//Proceedings of the 2017 IEEE International Conference on Computer Vision, Venice, Oct 22-29, 2017. Washington: IEEE Computer Society, 2017: 5908-5916.
[24] ZHANG H, XU T, LI H S, et al. StackGAN++: realistic image synthesis with stacked generative adversarial networks[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019, 41(8): 1947-1962.
[25] JOHNSON J, GUPTA A, LI F F. Image generation from scene graphs[C]//Proceedings of the 2018 IEEE Conference on Computer Vision and Pattern Recognition, Salt Lake City, Jun 18-22, 2018. Washington: IEEE Computer Society, 2018: 1219-1228.
[26] ZHU J Y, PAKR T, ISOLA P, et al. Unpaired image-to-image translation using cycle-consistent adversarial networks[C]//Proceedings of the 2017 IEEE International Conference on Computer Vision, Venice, Oct 22-29, 2017. Washington: IEEE Computer Society, 2017: 2242-2251.
[27] ISOLA P, ZHU J Y, ZHOU T H, et al. Image-to-image transla-tion with conditional adversarial networks[C]//Proceedings of the 2017 IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, Jul 21-26, 2017. Washington: IEEE Computer Society, 2017: 5967-5976.
[28] ZHU J Y, ZHANG R, PATHAK D, et al. Toward multimodal image-to-image translation[J]. arXiv:1711.11586, 2017.
[29] CHOI Y, CHOI M J, KIM M, et al. StarGAN: unified gene-rative adversarial networks for multi-domain image-to-image translation[J]. arXiv:1711.09020, 2017.
[30] BANSAL A, MA S G, RAMANAN D, et al. Recycle-GAN: unsupervised video retargeting[J]. arXiv:1808.05174, 2018.
[31] ZHANG H, GOODFELLOW I J, METAXAS D N, et al. Self-attention generative adversarial networks[J]. arXiv:1805. 08318v2, 2018.
[32] BROCK A, DONAHUE J, SIMONYAN K. Large scale GAN training for high fidelity natural image synthesis[J].arXiv:1809.11096, 2018.
[33] DONAHUE J, KR?HENBüHL P, DARRELL T, et al. Adver-sarial feature learning[J]. arXiv:1605.09782, 2016.
[34] DONAHUE J, SIMONYAN K. Large scale adversarial repre-sentation learning[J]. arXiv:1907.02544, 2019.
[35] KARRAS T, LAINE S, AILA T. A style-based generator architecture for generative adversarial networks[C]//Procee-dings of the 2019 IEEE Conference on Computer Vision and Pattern Recognition, Long Beach, Jun 16-20, 2019. Piscata-way: IEEE, 2019: 4401-4410.
[36] MAO X D, LI Q, XIE H R, et al. Least squares generative adversarial networks[C]//Proceedings of the 2017 IEEE Inter-national Conference on Computer Vision, Venice, Oct 22-29, 2017. Washington: IEEE Computer Society, 2017: 2813-2821.
[37] ARJOVSKY M, BOTTOU L. Towards principled methods for training generative adversarial networks[J]. arXiv:1701. 04862, 2017.
[38] BERTHELOT D, SCHUMM T, METZ L. BEGAN: boun-dary equilibrium generative adversarial networks[J]. arXiv:1703.10717, 2017.
[39] GULRAJANI I, AHMED F, ARJOVSKY M, et al. Improved training of Wasserstein GANs[J]. arXiv:1704.00028, 2017.
[40] NOWOZIN S, CSEKE B, TOMIOKA R. f-GAN: training generative neural samplers using variational divergence mini-mization[J]. arXiv:1606.00709, 2016.
[41] LEDIG C, THEIS L, HUSZAR F, et al. Photo-realistic single image super-resolution using a generative adversarial network[C]//Proceedings of the 2017 IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, Jul 21-26, 2017. Washington: IEEE Computer Society, 2017: 105-114.
[42] WANG X T, YU K, WU S X, et al. ESRGAN: enhanced superresolution generative adversarial networks[J]. arXiv:1809.00219, 2018.
[43] VOLKHONSKIY D, NAZAROV I, BORISENKO B. Stega-nographic generative adversarial networks[C]//Proceedings of the 12th International Conference on Machine Vision, Amsterdam, Nov 16-18, 2019. San Francisco: SPIE, 2019: 114333M.
[44] SHI H C, DONG J, WANG W, et al. SSGAN: secure stega-nography based on generative adversarial networks[C]//LNCS 10735: Proceedings of the 18th Pacific-Rim Conference on Multimedia, Harbin, Sep 28-29, 2017. Cham: Springer, 2017: 534-544.
[45] QIAN Y L, DONG J, WANG W, et al. Deep learning for steganalysis via convolutional neural networks[C]//Procee-dings of SPIE-The International Society for Optical Engi-neering. San Francisco: SPIE, 2015: 1-10.
[46] WANG Y J, NIU K, YANG X Y. Information hiding scheme based on generative adversarial network[J]. Journal of Com-puter Applications, 2018, 38(10): 2923-2928.
王耀杰, 钮可, 杨晓元. 基于生成对抗网络的信息隐藏方案[J]. 计算机应用, 2018, 38(10): 2923-2928.
[47] SANTANA E, HOTZ G. Learning a driving simulator[J]. arXiv:1608.01230, 2016．
[48] HUANG R, ZHANG S, LI T Y, et al. Beyond face rotation: global and local perception GAN for photorealistic and iden-tity preserving frontal view synthesis[J]. arXiv:1704.04086, 2017.
[49] REED S E, AKATA Z, YAN X C, et al. Generative adver-sarial text to image synthesis[J]. arXiv:1605.05396v1, 2016.
[50] ZHAO S Y, LI J W. Generative adversarial network for generating low-rank images[J]. Acta Automatica Sinica, 2018, 44(5): 829-839.
赵树阳, 李建武. 基于生成对抗网络的低秩图像生成方法[J]. 自动化学报, 2018, 44(5): 829-839.
[51] ZHU J Y, KR?HENBüHL P, SHECHTMAN E, et al. Gene-rative visual manipulation on the natural image manifold[C]//LNCS 9909: Proceedings of the 14th European Conference on Computer Vision, Amsterdam, Oct 11-14, 2016. Cham: Springer, 2016: 597-613.
[52] LI Y J, LIU S F, YANG J M, et al. Generative face comple-tion[J]. arXiv:1704.05838, 2017.
[53] LIU G L, REDA F A, SHIH K J, et al. Image inpainting for irregular holes using partial convolutions[C]//LNCS 11215: Proceedings of the 15th European Conference on Computer Vision, Munich, Sep 8-14, 2018. Cham: Springer, 2018: 89-105.
[54] CHEN J, DONG X L, LIANG J X, et al. Research on the local style transfer of clothing by CycleGAN based on atten-tion mechanism[J/OL]. Computer Engineering[2020-12-14]. https://doi.org/10.19678/j.issn.1000-3428.0059665.
陈佳, 董学良, 梁金星, 等. 基于注意力机制的CycleGAN服装局部风格迁移研究[J/OL]. 计算机工程[2020-12-14]. https://doi.org/10.19678/j.issn.1000-3428.0059665.
[55] LI C, ZHANG Y, HUANG C H. Improved super-resolution reconstruction of image based on generative adversarial net-works[J]. Computer Engineering and Applications, 2020, 56(4): 191-196.
李诚, 张羽, 黄初华. 改进的生成对抗网络图像超分辨率重建[J]. 计算机工程与应用, 2020, 56(4): 191-196.
[56] MA S, FU J L, CHEN C W, et al. DA-GAN: instance-level image translation by deep attention generative adversarial networks[C]//Proceedings of the 2018 IEEE Conference on Computer Vision and Pattern Recognition, Salt Lake City, Jun 18-22, 2018. Washington: IEEE Computer Society, 2018: 5657-5666.
[57] LIN Z F, YIN M X, YANG F, et al. Survey of image transla-tion based on conditional generative adversarial network[J]. Journal of Chinese Computer Systems, 2020, 41(12): 2569-2581.
林振峰, 尹梦晓, 杨锋, 等. 基于条件生成式对抗网络的图像转换综述[J]. 小型微型计算机系统, 2020, 41(12): 2569-2581.
[58] LIANG X D, LEE L, DAI W, et al. Dual motion GAN for future-flow embedded video prediction[J]. arXiv:1708.00284, 2017.
[59] TULYAKOV S, LIU M Y, YANG X D, et al. MoCoGAN: decomposing motion and content for video generation[C]//Proceedings of the 2018 IEEE Conference on Computer Vision and Pattern Recognition, Salt Lake City, Jun 18-22, 2018. Washington: IEEE Computer Society, 2018: 1526-1535.
[60] ZHANG Z Z, YANG L, ZHENG Y F. Translating and seg-menting multimodal medical volumes with cycle-and shape-consistency generative adversarial network[C]//Proceedings of the 2018 IEEE Conference on Computer Vision and Pattern Recognition, Salt Lake City, Jun 18-22, 2018. Washington: IEEE Computer Society, 2018: 9242-9251.
[61] CHEN J W, CHAO H Y, YANG M. Image blind denoising with generative adversarial network based noise modeling[C]//Proceedings of the 2018 IEEE Conference on Computer Vision and Pattern Recognition, Salt Lake City, Jun 18-22, 2018. Washington: IEEE Computer Society, 2018: 3155-3164.
[62] CHENG Z X, SUN H M, MASARU T, et al. Performance comparison of convolutional autoencoders, generative adver-sarial networks and super-resolution for image compression[C]//Proceedings of the 2018 IEEE Conference on Computer Vision and Pattern Recognition Workshops, Salt Lake City, Jun 18-22, 2018. Washington: IEEE Computer Society, 2018: 2613-2616.
[63] GADELHA M, MAJI S, WANG R. 3D shape induction from 2D views of multiple objects[J]. arXiv:1612.05872, 2016.
[64] LI Q Z, BAI W X, NIU J. Underwater image color correc-tion and enhancement based on improved cycle-consistent generative adversarial networks[J/OL]. Acta Automatica Sinica[2020-12-14]. https://doi.org/10.16383/j.aas.c200510.
李庆忠, 白文秀, 牛炯. 基于改进CycleGAN的水下图像颜色校正与增强[J/OL]. 自动化学报[2020-12-14]. https://doi.org/10.16383/j.aas.c200510.
[65] SUN X, DING X L. Data augmentation method based on generative adversarial networks for facial expression recogni-tion sets[J]. Computer Engineering and Applications, 2020, 56(4): 115-121.
孙晓, 丁小龙. 基于生成对抗网络的人脸表情数据增强方法[J]. 计算机工程与应用, 2020, 56(4): 115-121.
[66] YU H Y, LI G R, SU L, et al. Conditional GAN based indivi-dual and global motion fusion for multiple object tracking in UAV videos[J]. Pattern Recognition Letters, 2020, 131: 219-226.
[67] GAO L L, CHEN D Y, ZHAO Z, et al. Lightweight dyna-mic conditional GAN with pyramid attention for text-to-image synthesis[J]. Pattern Recognition, 202, 110: 107384.
[68] SALIMANS T, GOODFELLOW I J, ZAREMBA W, et al. Improved techniques for training GANs[C]//Proceedings of the Annual Conference on Neural Information Processing Systems 2016, Barcelona, Dec 5-10, 2016. Red Hook: Curran Associates, 2016: 2226-2234.
[69] HEUSEL M, RAMSAUER H, UNTERTHINER T, et al. GANs trained by a two time-scale update rule converge to a local Nash equilibrium[C]//Proceedings of the Annual Con-ference on Neural Information Processing Systems 2017, Long Beach, Dec 4-9, 2017. Red Hook: Curran Associates, 2017: 6626-6637.
[70] CHE T, LI Y R, JACOB A P, et al. Mode regularized genera-tive adversarial networks[J]. arXiv:1612.02136v5, 2016.
[71] GURUMURTHY S, SARVADEVABHATLA R K, BABU R V. DeLiGAN: generative adversarial networks for diverse and limited data[C]//Proceedings of the 2017 IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, Jul 21-26, 2017. Washington: IEEE Computer Society, 2017: 4941-4949.
[72] DZIUGAITE G K, ROY D M, GHAHRAMANI Z. Training generative neural networks via maximum mean discrepancy optimization[J]. arXiv:1505.03906, 2015.
[73] VALLENDER S S. Calculation of the Wasserstein distance between probability distributions on the line[J]. Theory of Probability & Its Applications, 1974, 18(4): 784-786.
[74] WANG Z, BOVIK A C, SHEIKH H R, et al. Image quality assessment: from error visibility to structural similarity[J]. IEEE Transactions on Image Processing, 2004, 13(4): 600-612.
[75] ZHANG L, ZHAO J Y, YE X L, et al. Collaborative genera-tion of confrontation networks[J]. Acta Automatica Sinica, 2018, 44(5): 804-810.
张龙, 赵杰煜, 叶绪伦, 等. 协作式生成对抗网络[J]. 自动化学报, 2018, 44(5): 804-810.
[76] LUCIC K K, MARCIN M, OLIVIER B, et al. Are GANs created equal? A large-scale study[J]. arXiv:1711.10337v1, 2017.
[77] ZHOU B L, ADETYA K, AGATA L, et al. Learning deep features for discriminative localization[J]. arXiv:1512.04150, 2015.
[78] JOLICOEUR M A. The relativistic discriminator: a key element missing from standard GAN[J]. arXiv:1807.00734, 2018.
[79] LI D, CHEN D C, JIN B H, et al. MAD-GAN: multivariate anomaly detection for time series data with generative adver-sarial networks[C]//LNCS 11730: Proceedings of the 28th International Conference on Artificial Neural Networks, Munich, Sep 17-19, 2019. Cham: Springer, 2019: 703-716.
[80] KARRAS T, AILA T, LAINE S, et al. Progressive growing of GANs for improved quality, stability, and variation[J]. arXiv:1710.10196v3, 2017.
[81] METZ L, POOLE B, PFAU D, et al. Unrolled generative adversarial networks[J]. arXiv:1611.02163, 2016.
[82] LIN Z N, KHETAN A, FANTI G C, et al. PacGAN: the power of two samples in generative adversarial networks[C]//Proceedings of the Annual Conference on Neural Informa-tion Processing Systems 2018, Montréal, Dec 3-8, 2018. Red Hook: Curran Associates, 2018: 1505-1514.
[83] KODALI N, ABERNETHY J, HAYS J, et al. On convergence and stability of GANs[J]. arXiv:1705.07215, 2017.
[84] NGUYEN T D, LE T, VU H, et al. Dual discriminator generative adversarial nets[J]. arXiv:1709.03831, 2017.