Semi-supervised self-training image classification method based on gamma divergence

doi:10.7523/j.ucas.2024.042

Abstract

Abstract:

In recent years， various semi-supervised self-training methods have made significant progress in solving image classification problems and have attracted widespread attention. However， even the state-of-the-art methods like FixMatch can suffer from training instability and extreme performance imbalance between classes due to the accumulation of pseudo-labeling errors. To address this issue， this paper introduces the concept of robust gamma divergence on top of the FixMatch method and proposes a semi-supervised self-training approach based on robust gamma divergence. In this method， we incorporate gamma divergence as a regularization term to the loss function of the training model to correct the mislabeled instances caused by error accumulation. Additionally， we conduct comparative experiments on artificially polluted simulated datasets and the CIFAR-10 dataset to demonstrate the superiority of the GammaFixMatch method in handling the problem of pseudo-label error accumulation.

Key words: image classification, semi-supervised, self-training, gamma divergence, deep learning, FixMatch

CLC Number:

O212

Ziqian MAO, Sanguo ZHANG. Semi-supervised self-training image classification method based on gamma divergence[J]. Journal of University of Chinese Academy of Sciences, 2026, 43(4): 433-443.

Figures/Tables 5

Fig.1 Impact of hyperparameter λ on performance over CIFAR-10 dataset

Fig.2 Pseudo-label error rate of the three experimental groups

Table 1 Total sample numbers and image sizes of two data sets

数据集	样本总数	类别数目	图像大小
CIFAR-10	60 000	10	$32 × 32$
CIFAR-100	60 000	100	$32 × 32$

Table 1 Total sample numbers and image sizes of two data sets

数据集	样本总数	类别数目	图像大小
CIFAR-10	60 000	10	$32 × 32$
CIFAR-100	60 000	100	$32 × 32$

Table 2 Number of labeled samples sampled in the CIFAR-10 dataset

实验组别

每类有标签

样本数

有标签

样本总数

有标签样本

占比

6.67 × 10 - 4

250

4.17 × 10 - 3

400

4000

6.67 × 10 - 2

Table 2 Number of labeled samples sampled in the CIFAR-10 dataset

实验组别

每类有标签

样本数

有标签

样本总数

有标签样本

占比

6.67 × 10 - 4

250

4.17 × 10 - 3

400

4000

6.67 × 10 - 2

Table 3 Error rates for CIFAR-10 and CIFAR-100 datasets

方法	CIFAR-10			CIFAR-100
方法	40个标签	250个标签	4 000个标签	400个标签
ПI-Model	-	54.26±3.97	14.01±0.38	-
Pseudo-Labeling	-	49.78±0.43	16.09±0.28	-
Mean Teacher	-	32.32±2.30	9.19±0.19	-
MixMatch	47.54±11.50	11.05±0.86	6.42±0.10	67.61±1.32
UDA	29.05±5.93	8.82±1.08	4.88±0.18	59.28±0.88
ReMixMatch	19.10±9.64	5.44±0.05	4.72±0.13	44.28±2.06
FixMatch（RA）	13.81±3.37	5.07±0.65	4.26±0.05	48.85±1.75
FixMatch（CTA）	11.39±3.35	5.07±0.33	4.31±0.15	49.95±3.01
GammaFixMatch	8.22 ± 2.81	4.82± 0.25	4.14± 0.04	47.40 ± 0.97

References 22

[1]	Krizhevsky A， Sutskever I， Hinton G E. ImageNet classification with deep convolutional neural networks［J］. Communications of the ACM， 2017， 60（6）： 84-90.DOI：10.1145/3065386 .
[2]	Grandvalet Y， Bengio Y. Semi-supervised learning by entropy minimization［C］//Proceedings of the 17th International Conference on Neural Information Processing Systems. December 13-18， 2004， Vancouver， BC， Canada. ACM， 2004： 529 - 536.
[3]	Chen T， Kornblith S， Swersky K， et al. Big self-supervised models are strong semi-supervised learners［C］//Proceedings of the 34th International Conference on Neural Information Processing Systems. December 6 - 12， 2020， Vancouver， BC， Canada. ACM， 2020： 22243 -22255. DOI： 10.5555/3495724.3497589 .
[4]	Lee D H. Pseudo-label： the simple and efficient semi-supervised learning method for deep neural networks［C］//Workshop on challenges in representation learning， ICML. 2013， 3（2）： 896.
[5]	Sohn K， Berthelot D， Li C L， et al. FixMatch： simplifying semi-supervised learning with consistency and confidence［C］//Proceedings of the 34th International Conference on Neural Information Processing Systems. December 6 - 12， 2020， Vancouver， BC， Canada. ACM， 2020： 596 - 608. DOI： 10.5555/3495724.3495775 .
[6]	Kirkpatrick J， Pascanu R， Rabinowitz N， et al. Overcoming catastrophic forgetting in neural networks［J］. Proceedings of the National Academy of Sciences of the United States of America， 2017， 114（13）： 3521-3526. DOI： 10.1073/pnas.1611835114 .
[7]	Chen B， Jiang J， Wang X， et al. Debiased self-training for semi-supervised learning［C］//Proceedings of the 36th International Conference on Neural Information Processing Systems. DOI：10.48550/2202.07136 .
[8]	Devlin J， Chang M W， Lee K， et al. BERT： pre-training of deep bidirectional transformers for language understanding［EB/OL］. arXiv 2018： 1810.04805. （2018-10-11）［2024-04-10］..
[9]	Zhou B L， Lapedriza A， Khosla A， et al. Places： a 10 million image database for scene recognition［J］. IEEE Transactions on Pattern Analysis and Machine Intelligence， 2018， 40（6）： 1452-1464. DOI：10.1109/TPAMI.2017.2723009 .
[10]	Fujisawa H， Eguchi S. Robust parameter estimation with a small bias against heavy contamination［J］. Journal of Multivariate Analysis， 2008， 99（9）： 2053-2081. DOI： 10.1016/j.jmva.2008.02.004 .
[11]	Ren M Y， Zhang S G， Ma S G， et al. Gene–environment interaction identification via penalized robust divergence［J］. Biometrical Journal. Biometrische Zeitschrift， 2022， 64（3）： 461-480. DOI： 10.1002/bimj.202000157 .
[12]	Hung H， Jou Z Y， Huang S Y. Robust mislabel logistic regression without modeling mislabel probabilities［J］. Biometrics， 2018， 74（1）： 145-154.DOI：10.1111/biom.12726 .
[13]	Jones M C， Hjort N L， Harris I R， et al. A comparison of related density‐based minimum divergence estimators［J］. Biometrika， 2001， 88（3）： 865-873. DOI： 10.1093/biomet/88.3.865 .
[14]	Zhang Z L， Sabuncu M R. Generalized cross entropy loss for training deep neural networks with noisy labels［C］//Proceedings of the 32nd International Conference on Neural Information Processing Systems. December 3 - 8， 2018， Montréal， Canada. ACM， 2018： 8792–8802. DOI： 10.5555/3327546.3327555 .
[15]	Amini M R， Feofanov V， Pauletto L， et al. Self-training： a survey［EB/OL］. arXiv 2022： 2202.12040.（2022-02-24）［2024-04-10］..
[16]	Zou Y， Yu Z D， Liu X F， et al. Confidence regularized self-training［C］//2019 IEEE/CVF International Conference on Computer Vision （ICCV）. Seoul， Korea （South）. IEEE， 2019： 5981-5990. DOI： 10.1109/ICCV.2019.00608 .
[17]	Huber P J. Robust statistics［M］. Wiley， 1982. DOI： 10.1002/0471725250 .
[18]	Ouali Y， Hudelot C， Tami M. Semi-supervised semantic segmentation with cross-consistency training［C］//2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition （CVPR）. Seattle， WA， USA. IEEE， 2020： 12671-12681. DOI： 10.1109/CVPR42600.2020.01269 .
[19]	Lai X， Tian Z T， Jiang L， et al. Semi-supervised semantic segmentation with directional context-aware consistency［C］//2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition （CVPR）. Nashville， TN， USA. IEEE， 2021： 1205-1214. DOI： 10.1109/CVPR46437.2021.00126 .
[20]	Cubuk E D， Zoph B， Shlens J， et al. Randaugment： practical automated data augmentation with a reduced search space［C］//2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops （CVPRW）. Seattle， WA， USA. IEEE， 2020： 3008-3017. DOI： 10.1109/CVPRW50498.2020.00359 .
[21]	Berthelot D， Carlini N， Cubuk E D， et al. ReMixMatch： semi-supervised learning with distribution alignment and augmentation anchoring［EB/OL］. arXiv 2019： 1911.09785.（2019-11-21）［2024-04-10］. .
[22]	Berthelot D， Carlini N， Goodfellow I， et al. MixMatch： a holistic approach to semi-supervised learning［EB/OL］. arXiv 2019： 1905.02249.（2019-05-06）［2024-04-10］..

Related Articles 15

[1]	Yucong WANG, Yong LI, Zhenqin YE, Boyu HU. Classification of insulator defect images with small samples based on meta learning and metric learning [J]. Journal of University of Chinese Academy of Sciences, 2026, 43(3): 422-431.
[2]	Cui JIANG, Jianbin JIAO. Knowledge-infused deep learning algorithm for vehicle trajectory prediction [J]. Journal of University of Chinese Academy of Sciences, 2026, 43(2): 209-217.
[3]	Ruiqi SUN, Wenjuan ZHANG, Zhen LI, Xuesong MA, Junlin MEI. Effect of patch size on super-resolution of large scene remote sensing images [J]. Journal of University of Chinese Academy of Sciences, 2026, 43(1): 93-103.
[4]	Huizhen WU, Sanguo ZHANG. Robust semi-supervised learning model based on model averaging and $γ$ -divergence [J]. Journal of University of Chinese Academy of Sciences, 2026, 43(1): 14-22.
[5]	YANG Tao, WANG Gaihua. Dual networks with hierarchical attention for fine-grained image classification [J]. Journal of University of Chinese Academy of Sciences, 2025, 42(6): 806-813.
[6]	ZHOU Qingze, GUO Qing. Multispectral remote sensing image pan-sharpening method based on multi-residual network [J]. Journal of University of Chinese Academy of Sciences, 2025, 42(4): 565-575.
[7]	ZENG Jianshun, LYU Yanjie, QIN Yuchu. Multi-scale semantic prior features guided street view image inpainting algorithm [J]. Journal of University of Chinese Academy of Sciences, 2025, 42(4): 496-507.
[8]	ZHONG Jinyan, CHEN Jun, LI Yu, WU Yewei, GE Xiaoqing. Cross-modal retrieval method based on MFF-SFE for remote sensing image-text [J]. Journal of University of Chinese Academy of Sciences, 2025, 42(2): 236-247.
[9]	ZHOU Wenxue, ZHANG Huachun. Lightweight network for fast ship detection in SAR images [J]. Journal of University of Chinese Academy of Sciences, 2024, 41(6): 776-785.
[10]	YU Yongsheng, LUO Tiejian. Seamless image completion via GAN inversion [J]. Journal of University of Chinese Academy of Sciences, 2024, 41(5): 705-714.
[11]	CHEN Jingwei, LI Yu, CHEN Jun, ZHANG Hongqun. Building extraction method based on MFF-Deeplabv3+ network for high-resolution remote sensing images [J]. Journal of University of Chinese Academy of Sciences, 2024, 41(5): 654-664.
[12]	HUANG Yulin, LIANG Lei, LI Weijun, XI Xiaohuan. Point cloud compression of deep learning based on multi-scale feature and attention mechanism [J]. Journal of University of Chinese Academy of Sciences, 2024, 41(5): 687-694.
[13]	LIN Yutong, WANG Hong, CHAI Tuanyao. Multi-scale featured convolution neural network-based soybean phenotypic prediction [J]. Journal of University of Chinese Academy of Sciences, 2024, 41(4): 468-476.
[14]	XIAO Jun, SHI Guangtian. Three-dimensional point cloud denoising [J]. Journal of University of Chinese Academy of Sciences, 2023, 40(5): 577-595.
[15]	YAO Mufeng, ZAN Luyang, LI Baipeng, LI Qingting, CHEN Zhengchao. Building change detection from remote sensing images using CAR-Siamese net [J]. Journal of University of Chinese Academy of Sciences, 2023, 40(3): 380-387.