显式鉴别驱动-未知类自动聚类的开放世界半监督学习方法

宋佳伦; 杜兰; 陈健

doi:10.11999/JEIT251291

显式鉴别驱动-未知类自动聚类的开放世界半监督学习方法

doi: 10.11999/JEIT251291 cstr: 32379.14.JEIT251291

西安电子科技大学雷达信号处理全国重点实验室西安 710071

基金项目: 国家自然科学基金(U24B20137, U21B2039, 62201433)，装备预研教育部联合基金(8091B03032401)，中央高校基金本科研业务费专项资金(QTZX23067)

详细信息

作者简介:
宋佳伦：男，博士生，研究方向为开放条件目标识别、增量学习、小样本学习

杜兰：女，博士，教授，博士生导师，研究方向为雷达目标识别、雷达信号处理、机器学习

陈健：男，博士，副教授，硕士生导师，研究方向为雷达目标检测识别与机器学习

通讯作者:
杜兰　dulan@mail.xidian.edu.cn

中图分类号: TN911.7; TP181
计量
- 文章访问数: 233
- HTML全文浏览量: 138
- PDF下载量: 8
- 被引次数: 0
出版历程
- 收稿日期: 2025-12-08
- 修回日期: 2026-06-22
- 录用日期: 2026-06-24
- 网络出版日期: 2026-07-04

Explicit Discrimination-driven Automatic Unknown Class Clustering for Open-World Semi-Supervised Learning

National Key Laboratory of Radar Signal Processing, Xidian University, Xi’an 710071, China

Funds: The National Natural Science Foundation of China (U24B20137, U21B2039, 62201433), The Equipment Preresearch Joint Fund of the Ministry of Education (8091B03032401), The Fundamental Research Funds for the Central Universities (QTZX23067)

摘要

摘要: 传统目标识别常基于闭集假设，即测试目标类别均属训练集。然而真实世界具有开放性，除训练集已知类外，识别任务还关注未知新目标的解译，要求模型兼具已知类识别与未知类发现并聚类的能力。针对上述问题，该文提出一种显式鉴别驱动-未知类自动聚类的直推式开放世界半监督学习(OWSSL)方法，综合利用少量标记已知类训练样本与大量无标记待测已知类/未知类样本学习模型，实现已知类识别与未知类聚类。所提方法重点包含基于已知类边界极值分布动态扩展的无标记已知类-未知类鉴别模块，结合极值理论建模标记已知类边界分布，并在半监督学习中对该分布进行动态完善，提升模型鉴别能力；对所鉴别高置信未知类样本，基于近邻交并比关系合并的未知类自动聚类模块进行近邻聚类簇合并，实现未知类自动聚类。鉴别模块和未知类自动聚类模块迭代优化。基于光学CIFAR-10与实测雷达数据的实验表明所提方法具有良好的已知类识别与未知类聚类性能。
- 目标识别 /
- 开放世界半监督学习 /
- 未知类鉴别 /
- 自动未知类聚类 /
- 极值理论
Abstract: Objective Traditional target recognition is generally developed under the closed-set assumption, in which all test classes are assumed to be included in the training set. In real-world scenarios, however, unlabeled unknown classes are commonly encountered. Therefore, recognition systems should simultaneously recognize known classes and discover and cluster unknown classes. To address this challenge, a transductive Open-World Semi-Supervised Learning (OWSSL) method driven by explicit discrimination and automatic unknown-class clustering is proposed. Methods The proposed method is trained using a small set of labeled known-class samples and a large set of unlabeled test samples containing both known and unknown classes. It consists of two complementary modules. The Dynamic Known-Unknown Class Discrimination (DKUCD) module models the known-class boundary distribution using Extreme Value Theory (EVT) and progressively refines the distribution with high-confidence known-class samples during semi-supervised learning to improve known-unknown discrimination. The Neighbor Intersection-Over-Union Cluster Merging (NIOUCM) module automatically clusters high-confidence unknown-class samples by merging neighboring clusters according to their intersection-over-union relationships. The DKUCD and NIOUCM modules are optimized iteratively to improve discrimination and unknown-class clustering jointly. Results and Discussions Experiments conducted on the optical CIFAR-10 dataset and measured radar datasets demonstrate that the proposed method achieves accurate known-class recognition while effectively clustering unknown classes. Conclusions By explicitly discriminating between known and unknown classes and automatically estimating unknown-class clusters, the proposed method improves both known-class recognition and unknown-class clustering under open-world conditions.
- Target recognition /
- Open-world semi-supervised learning /
- Unknown class discrimination /
- Automatic unknown-class clustering /
- Extreme Value Theory

HTML全文

图 1 直推式OWSSL示意图

下载: 全尺寸图片幻灯片

图 2 显式鉴别驱动-未知类自动聚类的OWSSL方法示意图

下载: 全尺寸图片幻灯片

图 3 EVT拟合示意图

下载: 全尺寸图片幻灯片

图 4 基于近邻交并比关系合并的未知类自动聚类模块示意图

下载: 全尺寸图片幻灯片

图 5 所提方法与TRSSL(对比方法中整体性能最优)的无标记已知类/未知类特征t-SNE

下载: 全尺寸图片幻灯片

图 6 所提方法与对比方法在50%, 100%已知类标注比例下的已知类识别/未知类聚类精度(未知类个数已知条件下)

下载: 全尺寸图片幻灯片

图 7 所提方法与对比方法(MSTAR, ATRNet-STAR对比中分别选择性能最优的TRSSL, LeGoGCD)的无标记已知类/未知类特征t-SNE

下载: 全尺寸图片幻灯片

图 8 所提无标记已知类-未知类鉴别模块有效性验证示意图(结果来自CIFAR-10数据)

下载: 全尺寸图片幻灯片

图 9 所提未知类自动聚类模块作用下的特征可视化示意图(其中星型表示类原型)(结果来自CIFAR-10数据)

下载: 全尺寸图片幻灯片

图 10 模型性能随超参数$ {\tau }_{1},\;{\tau }_{\text{c}},\;{{U}}_{0} $的变化曲线

下载: 全尺寸图片幻灯片

表 1 不同方法在CIFAR-10数据集上的RAUS(%), $ {\text{ACC}}_{\text{known}} $(%), $ {\text{ACC}}_{\text{cluster}} $(%)及$ {K+U} $估计个数

	方法	RAUS(%)		$ {\text{ACC}}_{\text{known}} $(%)		$ {\text{ACC}}_{\text{cluster}} $(%)		$ {K+U} $
无监督聚类	BootSC	-	-	87.70	76.07	88.92	77.42	[8,8,9]
无监督聚类	SNSCC	-	-	88.24	78.10	89.66	78.35	[8,8,9]
开放世界半监督学习	ORCA	91.71	82.45	88.20	80.02	90.40	81.22	[9,11,11]
	OpenLDN	94.54	86.55	93.40	88.35	93.65	84.14	[9,11,11]
	TRSSL	93.20	87.25	95.15	89.07	92.61	84.67	[9,11,11]
	PKOSSL	94.00	87.11	94.90	89.10	91.80	85.00	[9,11,11]
	LeGoGCD	98.80	92.25	64.64	62.62	98.45	89.42	[10,11,11]
	AFGCD	98.04	91.88	87.85	83.72	97.68	88.31	[10,11,11]
	本文方法	97.25	95.33	96.12	93.86	96.40	92.74	[10,10,11]

下载: 导出CSV

表 2 不同方法在MSTAR, ATRNet-STAR数据集上的RAUS(%), $ {\text{ACC}}_{\text{known}} $(%), $ {\text{ACC}}_{\text{cluster}} $(%)及$ {K+U} $估计个数

	方法	RAUS		$ {\text{ACC}}_{\text{known}} $		$ {\text{ACC}}_{\text{cluster}} $		$ {K+U} $
	方法	MSTAR	ATRNet-STAR	MSTAR	ATRNet-STAR	MSTAR	ATRNet-STAR	MSTAR	ATRNet-STAR
无监督聚类	BootSC	-	-	72.80	68.62	61.18	25.99	[7,8,8]	[36,37,38]
无监督聚类	SNSCC	-	-	75.15	70.33	62.02	26.96	[7,8,8]	[36,37,38]
开放世界半监督学习	ORCA	70.03	54.68	80.73	83.76	67.99	31.70	[8,9,9]	[38,39,41]
	OpenLDN	72.32	58.53	83.09	88.12	70.66	34.24	[8,9,9]	[38,39,41]
	TRSSL	73.55	60.12	83.82	87.05	70.24	37.98	[8,9,9]	[38,39,41]
	PKOSSL	74.18	63.66	81.87	85.20	72.61	38.13	[8,9,9]	[38,39,41]
	LeGoGCD	80.90	75.08	60.95	89.17	78.45	42.22	[8,9,10]	[38,40,42]
	AFGCD	78.60	72.66	62.45	88.41	76.68	40.37	[8,9,10]	[38,40,42]
	本文方法	86.59	84.94	88.27	90.15	84.21	50.09	[10,10,10]	[39,39,40]

下载: 导出CSV

表 3 所提模块的消融实验结果RAUS, $ {\text{ACC}}_{\text{known}} $, $ {\text{ACC}}_{\text{cluster}} $

	无标记已知类-未知类鉴别模块	未知类自动聚类模块	RAUS		$ {\text{ACC}}_{\text{known}} $		$ {\text{ACC}}_{\text{cluster}} $
Baseline	×	×	87.25	60.12	89.07	87.05	83.67	37.98
所提模块	Π	×	95.09	80.97	91.54	88.60	86.35	44.85
	×	Π	86.76	62.79	90.19	87.33	89.01	39.01
	Π	Π	95.33	84.94	93.86	90.15	92.74	50.09

下载: 导出CSV

参考文献(16)

[1]	张东阳, 陆子轩, 刘军民, 等. 深度模型的持续学习综述: 理论、方法和应用[J]. 电子与信息学报, 2024, 46(10): 3849–3878. doi: 10.11999/JEIT240095. ZHANG Dongyang, LU Zixuan, LIU Junmin, et al. A survey of continual learning with deep networks: Theory, method and application[J]. Journal of Electronics & Information Technology, 2024, 46(10): 3849–3878. doi: 10.11999/JEIT240095.
[2]	CAO Kaidi, BRBIC M, and LESKOVEC J. Open-world semi-supervised learning[C]. International Conference on Learning Representations, 2022: 22837–22855.
[3]	RIZVE M N, KARDAN N, KHAN S, et al. Openldn: Learning to discover novel classes for open-world semi-supervised learning[C]. 17th European Conference on Computer Vision–ECCV 2022, Tel Aviv, Israel, 2022: 382–401. doi: 10.1007/978-3-031-19821-2_22.
[4]	RIZVE M N, KARDAN N, and SHAH M. Towards realistic semi-supervised learning[C]. 17th European Conference on Computer Vision - ECCV 2022, Tel Aviv, Israel, 2022: 437–455. doi: 10.1007/978-3-031-19821-2_25.
[5]	MARCO C. Sinkhorn distances: Lightspeed computation of optimal transport[C]. Advances in Neural Information Processing Systems, Stateline 26: 27th Annual Conference on Neural Information Processing Systems 2013, Lake Tahoe, United States, 2013: 2292–2300.
[6]	ZHAO Tianhao, LIN Yutian, WU Yu, et al. Promote knowledge mining towards open-world semi-supervised learning[J]. Pattern Recognition, 2024, 149: 110259. doi: 10.1016/j.patcog.2024.110259.
[7]	CAO Xinzi, ZHENG Xiawu, WANG Guanhong, et al. Solving the catastrophic forgetting problem in generalized category discovery[C]. The 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Seattle, USA, 2024, 16880–16889. doi: 10.1109/CVPR52733.2024.01597.
[8]	XU Qiyu, HU Zhanxuan, DUAN Yu, et al. A hidden stumbling block in generalized category discovery: Distracted attention[C]. The 2025 IEEE/CVF International Conference on Computer Vision (ICCV), Honolulu, USA, 2025: 405–414. doi: 10.1109/ICCV51701.2025.00045.
[9]	KRIZHEVSKY A and HINTON G E. Learning multiple layers of features from tiny images[R]. 2009. .
[10]	ROSS T D, WORRELL S W, VELTEN V J, et al. Standard SAR ATR evaluation experiments using the MSTAR public release data set[J]. Algorithms for Synthetic Aperture Radar Imagery V, 1998, 3370: 566–573. doi: 10.1117/12.321859.
[11]	LIU Yongxiang, LI Weijie, LIU Li, et al. ATRNet-STAR: A large dataset and benchmark towards remote sensing object recognition in the wild[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2026, 48(6): 6735–6753. doi: 10.1109/TPAMI.2026.3658649.
[12]	MOVSHOVITZ-ATTIAS Y, TOSHEV A, LEUNG T K, et al. No fuss distance metric learning using proxies[C]. The 2017 IEEE International Conference on Computer Vision (ICCV), Venice, Italy, 2017: 360–368. doi: 10.1109/ICCV.2017.47.
[13]	陈健, 雍奇锋, 杜兰, 等. 结合未知类特征生成与分类得分修正的SAR目标开集识别方法[J]. 电子与信息学报, 2024, 46(10): 3890–3907. doi: 10.11999/JEIT240138. CHEN Jian, YONG Qifeng, DU Lan, et al. An open set recognition method for SAR targets combining unknown feature generation and classification score modification[J]. Journal of Electronics & Information Technology, 2024, 46(10): 3890–3907. doi: 10.11999/JEIT240138.
[14]	杜兰, 李逸明, 薛世鲲, 等. 结合相似度预测和阈值自动求解的开集条件下毫米波雷达点云步态识别方法[J]. 电子与信息学报, 2025, 47(6): 1850–1863. doi: 10.11999/JEIT241034. DU Lan, LI Yiming, XUE Shikun, et al. Millimeter-wave radar point cloud gait recognition method under open-set conditions based on similarity prediction and automatic threshold estimation[J]. Journal of Electronics & Information Technology, 2025, 47(6): 1850–1863. doi: 10.11999/JEIT241034.
[15]	GUO Wengang, YE Wei, CHEN Chunchun, et al. Bootstrap deep spectral clustering with optimal transport[J]. IEEE Transactions on Multimedia, 2026, 28: 531–544. doi: 10.1109/TMM.2025.3623492.
[16]	DUAN Yu, CHEN Huimin, ZHANG Runxin, et al. Soft neighbors supported contrastive clustering[J]. IEEE Transactions on Image Processing, 2025, 34: 4315–4327. doi: 10.1109/TIP.2025.3583194.