基于条件生成对抗和泊松融合的图像数据增强方法

doi:10.3785/j.issn.1008-973X.2026.12.005

浙江大学学报(工学版)

2026, Vol. 60

Issue (12): 2592-2601 DOI: 10.3785/j.issn.1008-973X.2026.12.005

计算机技术

基于条件生成对抗和泊松融合的图像数据增强方法

陈可嘉(

),曾威文,王朕卿

福州大学经济与管理学院，福建福州 350108

Image data enhancement method based on conditional generative adversarial network and Poisson blending

Kejia CHEN(

),Weiwen ZENG,Zhenqing WANG

School of Economics and Management, Fuzhou University, Fuzhou 350108, China

全文: PDF(2625 KB) HTML

摘要：

为了解决图像数据集中长尾分布与标签稀缺问题，提出基于条件生成对抗学习和泊松融合的图像数据增强方法. 该方法通过引入条件信息构建具有梯度惩罚的Wasserstein条件生成对抗网络以实现类别可控的图像生成，采用泊松融合提升生成图像的平滑度与视觉连贯性. 在CIFAR-10和“华为云”垃圾分类数据集上的实验表明，该方法在生成质量和训练稳定性上优于StyleGAN、ACGAN、WGAN-Div等主流模型，且以较短的生成时间接近或超越去噪扩散概率模型的性能，通过消融实验验证了各模块的有效性. 该方法显著提升了生成图像的结构完整性与语义一致性，有效解决了数据不平衡问题，具有良好的实用价值.

关键词： 图像数据增强; 长尾分布; 标签稀缺; 条件生成对抗网络; 泊松融合

Abstract:

An image data augmentation method based on conditional generative adversarial learning and Poisson blending was proposed in order to address the long-tailed distribution and label scarcity in image datasets. A Wasserstein conditional generative adversarial network with gradient penalty was constructed by incorporating class-conditional information in order to enable class-controllable image generation. Poisson blending was employed to improve the smoothness and visual coherence of the generated images. Experiments conducted on CIFAR-10 and the Huawei-Garbage dataset demonstrated that the proposed method outperformed representative models such as StyleGAN, ACGAN and WGAN-Div in terms of generation quality and training stability. The proposed method approached or exceeded the performance of denoising diffusion probabilistic models with a shorter generation time. Ablation studies verified the effectiveness of each constituent module. The proposed method significantly improves the structural integrity and semantic consistency of generated images, effectively alleviates the data imbalance problem, and demonstrates strong practical applicability.

Key words: image data augmentation long-tailed distribution label scarcity conditional generative adversarial network Poisson blending

收稿日期: 2025-09-14 出版日期: 2026-10-21

CLC:

TP 391

基金资助: 国家自然科学基金资助项目(72374030)；福建省财政厅专项资金资助项目(0300-83022110).

作者简介: 陈可嘉（1978—），男，教授，博士，从事工业工程、系统工程研究. orcid.org/0000-0002-3693-808X. E-mail：kjchen@fzu.edu.cn

	服务
	把本文推荐给朋友
	加入引用管理器
	E-mail Alert
	作者相关文章
	陈可嘉
	曾威文
	王朕卿

引用本文:

陈可嘉,曾威文,王朕卿. 基于条件生成对抗和泊松融合的图像数据增强方法[J]. 浙江大学学报(工学版), 2026, 60(12): 2592-2601.

Kejia CHEN,Weiwen ZENG,Zhenqing WANG. Image data enhancement method based on conditional generative adversarial network and Poisson blending. Journal of ZheJiang University (Engineering Science), 2026, 60(12): 2592-2601.

链接本文:

https://www.zjujournals.com/eng/CN/10.3785/j.issn.1008-973X.2026.12.005 或 https://www.zjujournals.com/eng/CN/Y2026/V60/I12/2592

图 1 PB-WCGAN-GP网络的结构

图 2 PB-WCGAN-GP生成器的结构

图 3 CBAM的结构

图 4 PB-WGAN-GP评价器的结构

图 5 WGAN收敛判断的原理

图 6 PB融合去噪模块的结构

表 1 不同生成网络的参数量和FLOPs对比

表 2 不同生成网络的IS与FID对比

表 3 不同数据集上的消融实验结果

表 4 数据增强前、后的对比实验

图 7 在CIFAR-10上的生成图像对比

图 8 在Huawei-Garbage上的生成图像对比

图 9 PB-WCGAN-GP在CIFAR-10-cat上的专项图像生成

图 10 在Huawei-Garbage-20（“可回收物/快递纸袋”）上的专项图像生成对比

1	Javed H, El-sappagh S, Abuhmed T Robustness in deep learning models for medical diagnostics: security and adversarial challenges towards robust AI applications[J]. Artificial Intelligence Review, 2024, 58 (1): 12 doi: 10.1007/s10462-024-11005-9
2	Buda M, Maki A, Mazurowski M A A systematic study of the class imbalance problem in convolutional neural networks[J]. Neural Networks, 2018, 106: 249- 259 doi: 10.1016/j.neunet.2018.07.011
3	魏秀参长尾分布图像识别: 途径与进展[J]. 中国基础科学, 2023, 25 (5): 48- 55 Wei Xiucan Long-tailed image recognition: approaches and advances[J]. China Basic Science, 2023, 25 (5): 48- 55 doi: 10.3969/j.issn.1009-2412.2023.05.007
4	Yang Suorong, Yang Hongchao, Shen Furao, et al Image data augmentation for deep learning: a survey[J]. Journal of Software, 2025, 36 (3): 1390- 1412 doi: 10.13328/j.cnki.jos.007263
5	Goodfellow I, Pouget-Abadie J, Mirza M, et al Generative adversarial networks[J]. Communications of the ACM, 2020, 63 (11): 139- 144 doi: 10.1145/3422622
6	Lesmes-Leon D N, Dengel A, Ahmed S Systematic review of generative adversarial networks (GANs) in cell microscopy: trends, practices, and impact on image augmentation[J]. PLoS One, 2025, 20 (6): e0291217 doi: 10.1371/journal.pone.0291217
7	Saxena D, Cao J Generative adversarial networks (GANs): challenges, solutions, and future directions[J]. ACM Computing Surveys, 2022, 54 (3): 1- 42
8	Gulrajani I, Ahmed F, Arjovsky M, et al. Improved training of Wasserstein GANs [C]//Advances in Neural Information Processing Systems 30. Long Beach: Curran Associates, Inc., 2017: 5767–5777.
9	Yang X, He H, Li J, et al Toward optimal risk-averse configuration for HESS with CGANs-based PV scenario generation[J]. IEEE Transactions on Systems, Man, and Cybernetics: Systems, 2021, 51 (3): 1779- 1793 doi: 10.1109/TSMC.2019.2905776
10	Pérez P, Gangnet M, Blake A Poisson image editing[J]. ACM Transactions on Graphics, 2003, 22 (3): 313- 318 doi: 10.1145/882262.882269
11	Chen T, Ren J. Integrating GAN and texture synthesis for enhanced road damage detection [PP/OL]. arXiv(2023-09-13)[2025-09-01]. https://arxiv.org/abs/2309.06747.
12	He K, Zhang X, Ren S, et al. Delving deep into rectifiers: surpassing human-level performance on ImageNet classification [C]//Proceedings of the IEEE International Conference on Computer Vision. Santiago: IEEE, 2016: 1026–1034.
13	Krizhevsky A. Learning multiple layers of features from tiny images [EB/OL]. (2009-04-08)[2025-09-01]. https://www.cs.toronto.edu/~kriz/learning-features-2009-TR.pdf.
14	高静, 段中兴, 何宇超多尺度特征融合的轻量型垃圾分类方法[J]. 小型微型计算机系统, 2023, 44 (2): 376- 382 Gao Jing, Duan Zhongxing, He Yuchao Lightweight garbage classification method based on multi-scale feature fusion[J]. Journal of Chinese Computer Systems, 2023, 44 (2): 376- 382 doi: 10.20009/j.cnki.21-1106/TP.2021-0480
4	杨锁荣, 杨洪朝, 申富饶, 等面向深度学习的图像数据增强综述[J]. 软件学报, 2025, 36 (3): 1390- 1412 doi: 10.13328/j.cnki.jos.007263
15	Betzalel E, Penso C, Fetaya E Evaluation metrics for generative models: an empirical study[J]. Machine Learning and Knowledge Extraction, 2024, 6 (3): 1531- 1544 doi: 10.3390/make6030073
16	Heusel M, Ramsauer H, Unterthiner T, et al. GANs trained by a two time-scale update rule converge to a local Nash equilibrium [C]//Advances in Neural Information Processing Systems 30. Long Beach: Curran Associates, Inc., 2017: 6626–6637.
17	Petersen F, Kuehne H, Borgelt C, et al. Differentiable top-k classification learning [C]//Proceedings of the International Conference on Machine Learning. Baltimore: PMLR, 2022.
18	Christen P, Hand D J, Kirielle N A review of the F-measure: its history, properties, criticism, and alternatives[J]. ACM Computing Surveys, 2024, 56 (3): 73 doi: 10.1145/3606367
19	Xie S, Girshick R, Dollár P, et al. Aggregated residual transformations for deep neural networks [C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Honolulu: IEEE, 2017: 5987–5995.
20	Zhang X, Zhou X, Lin M, et al. ShuffleNet: an extremely efficient convolutional neural network for mobile devices [C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Salt Lake City: IEEE, 2018: 6848–6856.
21	Karras T, Laine S, Aila T. A style-based generator architecture for generative adversarial networks [C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Long Beach: IEEE, 2019: 4396–4405.
22	Odena A, Olah C, Shlens J. Conditional image synthesis with auxiliary classifier GANs [C]//Proceedings of the 34th International Conference on Machine Learning. Sydney: PMLR, 2017: 2642–2651.
23	Qu F, Liao H, Lu M, et al Recognition of drill string vibration state based on WGAN-div and CNN-IWPSO-SVM[J]. Geoenergy Science and Engineering, 2024, 243: 213342 doi: 10.1016/j.geoen.2024.213342
24	Ho J, Jain A, Abbeel P. Denoising diffusion probabilistic models [C]//Advances in Neural Information Processing Systems 33. Vancouver: Curran Associates, Inc., 2020: 6840–6851.
25	Huang Y, Gokaslan A, Kuleshov V, et al. The GAN is dead; long live the GAN! A Modern GAN Baseline [C]//Advances in Neural Information Processing Systems 37. Vancouver: Curran Associates, 2024: 44177–44215.
26	Liu H, Gu X, Samaras D. A two-step computation of the exact GAN Wasserstein distance [C]//Proceedings of the 35th International Conference on Machine Learning. Stockholm: PMLR, 2018: 3159–3168.

[1]	詹燕,胡蝶,汤洪涛,鲁建厦,谭健,刘长睿. 基于改进生成对抗网络的图像数据增强方法[J]. 浙江大学学报(工学版), 2023, 57(10): 1998-2010.
[2]	白文超,韩希先,王金宝. 基于条件生成模型的高效近似查询处理框架[J]. 浙江大学学报(工学版), 2022, 56(5): 995-1005.
[3]	陈雪云,黄小巧,谢丽. 基于多尺度条件生成对抗网络血细胞图像分类检测方法[J]. 浙江大学学报(工学版), 2021, 55(9): 1772-1781.

Viewed

Full text

Abstract

Cited

Shared

Discussed