基于自适应原型特征类矫正的小样本学习方法

赵红; 钟杨清; 金杰; 邹林华

doi:10.16383/j.aas.c240312

基于自适应原型特征类矫正的小样本学习方法

doi: 10.16383/j.aas.c240312

赵红^{1, 2,},
钟杨清^{1, 2,},
金杰^{1, 2,},
邹林华^{1, 2,}

1.
闽南师范大学计算机学院漳州 363000
2.
闽南师范大学数据科学与智能计算福建省高校重点实验室漳州 363000

基金项目: 国家自然科学基金 (62376114)资助

详细信息

作者简介:
赵红：闽南师范大学计算机学院教授. 2019年获得天津大学博士学位. 主要研究方向为粒计算, 小样本学习及分层分类. 本文通信作者. E-mail: hongzhaocn@163.com

钟杨清：闽南师范大学计算机学院硕士研究生. 主要研究方向为小样本学习. E-mail: yangqingzhong0@163.com

金杰：闽南师范大学计算机学院硕士研究生. 主要研究方向为机器学习和粒度计算中的少样本学习和分层分类应用. E-mail: kimjee@126.com

邹林华：闽南师范大学计算机学院硕士研究生. 主要研究方向为机器学习和粒度计算中的少样本学习和分层分类应用. E-mail: zlh1836065471@163.com

计量
- 文章访问数: 106
- HTML全文浏览量: 46
- 被引次数: 0
出版历程
- 收稿日期: 2024-06-03
- 录用日期: 2024-08-29
- 网络出版日期: 2024-09-22

Few-shot Learning Based on Class Rectification via Adaptive Prototype Features

ZHAO Hong^{1, 2
,},
ZHONG Yang-Qing^{1, 2
,},
JIN Jie^{1, 2
,},
ZOU Lin-Hua^{1, 2
,}

1.
School of Computer Science, Minnan Normal University, Zhangzhou 363000
2.
Key Laboratory of Data Science and Intelligent Application, Minnan Normal University, Zhangzhou 363000

Funds: Supported by National Natural Science Foundation of China (62376114)

More Information

Author Bio:
ZHAO Hong　Professor at the School of Computer Science, Minnan Normal University. She received her Ph.D. degree from Tianjin University in 2019. Her research interest covers granular computing, few-shot learning, and data mining for hierarchical classification. Corresponding author of this paper

ZHONG Yang-Qing　Master student at the School of Computer Science, Minnan Normal University. Her research interest covers few-shot learning

JIN Jie　Master student at the School of Computer Science, Minnan Normal University. His research interest covers few-shot learning and hierarchical classification application in machine learning and granular computing

ZOU Lin-Hua　Master student at the School of Computer Science, Minnan Normal University. His research interest covers few-shot learning and hierarchical classification application in machine learning and granular computing

摘要

摘要: 针对小样本学习过程上样本数量不足导致的性能下降问题, 基于原型网络的小样本学习方法通过实现查询样本与支持样本原型特征间的距离度量, 从而达到很好的分类性能. 然而, 这种方法直接将支持集样本均值视为类原型, 在一定程度上加剧了对样本数量稀少情况下的敏感性. 针对此问题, 提出了基于自适应原型特征类矫正的小样本学习方法(Few-shot learning based on class rectification via adaptive prototype features, CRAPF), 通过自适应生成原型特征来缓解模型对数据细微变化的过度响应, 并同步实现类边界的精细化调整. 首先, 使用卷积网络构建自适应原型特征生成模块, 该模块采用非线性映射获取更为稳健的原型特征, 有助于减弱异常值对原型构建的影响. 其次, 通过对原型生成过程的优化, 提升了不同类间原型表示的区分度, 进而强化了原型特征对于类别表征的整体效能. 最后, 在3个广泛使用的基准数据集上的实验结果显示, 该方法提升了小样本学习任务的表现. 例如, 在5类5样本设置下, CRAPF在MiniImageNet和CIFAR-FS上的准确率比其他模型至少提高了2.06% 和2.30%.
- 小样本学习 /
- 原型网络 /
- 原型特征 /
- 类矫正
Abstract: In response to the performance degradation issue from inadequate sample sizes during few-shot learning, prototypical network-based few-shot learning techniques achieve commendable classification capabilities by measuring the distance metrics between query sample features and the prototype features of support samples. However, this method directly treats the mean of the support set samples as class prototypes, exacerbating sensitivity to scarcity of samples. To counteract this problem, we propose few-shot learning based on class rectification via adaptive prototype features (CRAPF), which adaptively generate prototype features for each class. It mitigates the model's responsiveness to minor data variations and concurrently realizes class rectification. First, we construct an adaptive prototype feature generation module using convolutional neural networks. This module leverages nonlinear mappings to obtain adaptive prototype features, thereby mitigating the impact of outliers on prototype construction. Second, we optimize the prototype generation process, and the discriminative power among prototype representations across distinct classes is heightened. Finally, experiments conducted on three extensively utilized benchmark datasets reveal that this method significantly enhances the performance of few-shot learning tasks. For instance, under the 5-way-5-shot setting, CRAPF outperforms other models on MiniImageNet and CIFAR-FS with an accuracy improvement of at least 2.06% and 2.30%.
- Few-shot learning /
- prototype network /
- prototypical features /
- class rectification

HTML全文

图 1 固定原型和自适应原型生成示例 (5类5样本)

Fig. 1 Examples of fixed prototype and adaptive prototype generation (5-way-5-shot)

下载: 全尺寸图片幻灯片

图 2 基于5类5样本的模型框架图

Fig. 2 Model framework based on 5-way-5-shot

下载: 全尺寸图片幻灯片

图 3 自适应原型特征生成模块

Fig. 3 Adaptive prototype feature generation module

下载: 全尺寸图片幻灯片

图 4 数据集部分代表性图像

Fig. 4 Representative images in the datasets

下载: 全尺寸图片幻灯片

图 5 在不同数据集上自适应原型和预训练模块在CRAPF中的性能比较

Fig. 5 Performance comparison of adaptive prototype and pre-training module on different datasets in CRAPF

下载: 全尺寸图片幻灯片

图 6 应用不同N类K样本策略的性能比较

Fig. 6 Performance comparison using different N-way-K-shot strategies

下载: 全尺寸图片幻灯片

图 7 CRAPF的可视化

Fig. 7 Visualization of CRAPF

下载: 全尺寸图片幻灯片

表 1 实验数据集

Table 1 Experimental datasets

数据集	样本数量	训练/验证/测试类别数量	图片大小
MiniImageNet	60 000	64/16/20	84$ \times $84像素
CIFAR-FS	60 000	64/16/20	32$ \times $32像素
FC100	60 000	60/20/20	32$ \times $32像素

下载: 导出CSV

表 2 应用不同$ N $类$ K $样本策略的性能比较 (%)

Table 2 Performance comparison using different $ N $-way-$ K $-shot strategies (%)

数据集	$ N $类$ K $样本	ProtoNet	CRAPF
CIFAR-FS	5类5样本	75.68	85.11
	5类10样本	81.17	87.34
	10类5样本	68.18	74.30
FC100	5类5样本	50.96	54.56
	5类10样本	57.20	59.13
	10类5样本	36.69	37.89

下载: 导出CSV

表 3 应用不同网络的实验结果 (%)

Table 3 Experimental results using different networks (%)

网络	数据集	ProtoNet	CRAPF
ResNet12	MiniImageNet	72.08	77.08
	CIFAR-FS	75.68	85.11
	FC100	50.96	54.56
ResNet18	MiniImageNet	73.68	74.40
	CIFAR-FS	72.83	84.93
	FC100	47.50	53.33

下载: 导出CSV

表 4 MiniImageNet数据集上的对比实验结果 (%)

Table 4 Comparative experimental results on the MiniImageNet dataset(%)

模型	网络	5类1样本	5类5样本
TEAM	ResNet18	60.07	75.90
TransCNAPS	ResNet18	55.60	73.10
MTUNet	ResNet18	58.13	75.02
IPN	ResNet10	56.18	74.60
ProtoNet*	ResNet12	53.42	72.08
TADAM	ResNet12	58.50	76.70
SSR	ResNet12	68.10	76.90
MDM-Net	ResNet12	59.88	76.60
CRAPF*	ResNet12	59.38	77.08

下载: 导出CSV

表 5 CIFAR-FS数据集上的对比实验结果 (%)

Table 5 Comparative experimental results on the CIFAR-FS dataset (%)

模型	网络	5类1样本	5类5样本
ProtoNet*	ResNet12	56.86	75.68
Shot-Free	ResNet12	69.20	84.70
TEAM	ResNet12	70.40	80.30
DeepEMD	ResNet12	46.50	63.20
DSN	ResNet12	72.30	85.10
MTL	ResNet12	69.50	84.10
DSMNet	ResNet12	60.66	79.26
MTUNet	ResNet18	67.43	82.81
CRAPF*	ResNet12	72.34	85.11

下载: 导出CSV

表 6 FC100数据集上的对比实验结果 (%)

Table 6 Comparative experimental results on the FC100 dataset(%)

模型	网络	5类1样本	5类5样本
ProtoNet*	ResNet12	37.50	50.96
SimpleShot	ResNet10	40.13	53.63
Baseline2020	ResNet12	36.82	49.72
TADAM	ResNet12	40.10	56.10
CRAPF*	ResNet12	40.44	54.56

下载: 导出CSV

参考文献(42)

[1]	Jiang H, Diao Z, Shi T, Zhou Y, Wang F, Hu W, et al. A review of deep learning-based multiple-lesion recognition from medical images: classification, detection and segmentation. Computers in Biology and Medicine, 2023Article No. 106726
[2]	赵凯琳, 靳小龙, 王元卓. 小样本学习研究综述. 软件学报, 2021, 32(2): 349−369 Zhao Kai-Lin, Jin Xiao-Long, Wang Yuan-Zhuo. Survey on few-shot learning. Journal of Software, 2021, 32(2): 349−369
[3]	赵一铭, 王佩瑾, 刁文辉, 孙显, 邓波. 基于通道注意力机制的小样本SAR飞机图像分类方法. 南京大学学报(自然科学版), 2024, 60(03): 464−476 Zhao Yi-Ming, Wang Pei-Jin, Diao Wen-Hui, Sun Xian, Deng Bo. Few-shot SAR aircraft image classification method based on channel attention mechanism. Journal of Nanjing University (Natural Science), 2024, 60(03): 464−476
[4]	Li F F, Fergus R, Perona P. A Bayesian approach to unsupervised one-shot learning of object categories. In: Proceedings of the IEEE International Conference on Computer Vision. Nice, Franc: IEEE, 2003. 1134−1141
[5]	刘颖, 雷研博, 范九伦, 王富平, 公衍超, 田奇. 基于小样本学习的图像分类技术综述. 自动化学报, 2021, 47(2): 297−315 Liu Ying, Lei Yan-Bo, Fan Jiu-Lun, Wang Fu-Ping, Gong Yan-Chao, Tian Qi. Survey on image classification technology based on small sample learning. Acta Automatica Sinica, 2021, 47(2): 297−315
[6]	Li X, Yang X, Ma Z, Xue J. Deep metric learning for few-shot image classification: A review of recent developments. Pattern Recognition, 2023Article No. 109381
[7]	Snell J, Swersky K, Zemel R. Prototypical networks for few-shot learning. In: Proceedings of the 31th International Conference on Neural Information Processing Systems. Long Beach, USA: 2017. 4077−4087
[8]	Qiang W, Li J, Su B, Fu J, Xiong H, Wen J. Meta attention-generation network for cross-granularity few-shot learning. IEEE International Journal of Computer Vision, 2023, 131(5): 1211−1233 doi: 10.1007/s11263-023-01760-7
[9]	Li W, Wang L, Xu J, Huo J, Gao Y, Luo J. Revisiting local descriptor based image-to-class measure for few-shot learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Long Beach, USA: IEEE, 2019. 7260−7268
[10]	Xu W, Xu Y, Wang H, Tu Z. Attentional constellation nets for few-shot learning. In: Proceedings of the International Conference on Learning Representations. Vienna, Austria: ICLR, 2021. 1−12
[11]	Huang H, Zhang J, Yu L, Zhang J, Wu Q, Xu C. TOAN: target-oriented alignment network for fine-grained image categorization with few labeled samples. IEEE Transactions on Circuits and Systems for Video Technology, 2021, 32(2): 853−866
[12]	Xie J, Long F, Lv J, Wang Q, Li P. Joint distribution matters: deep brownian distance covariance for few-shot classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. New Orleans, USA: IEEE, 2022. 7972−7981
[13]	Luo X, Wei L, Wen L, Yang J, Xie L, Xu Z, et al. Rectifying the shortcut learning of background for few-shot learning. In: Proceedings of the 35th International Conference on Neural Information Processing Systems. Virtual Event: 2021. 13073−13085
[14]	Kwon H, Jeong S, Kim S, Sohn K. Dual prototypical contrastive learning for few-shot semantic segmentation. arXiv preprint arXiv: 2111.04982, 2021.
[15]	Li J, Liu G. Few-shot image classification via contrastive self-supervised learning. arXiv preprint arXiv: 2008.09942, 2020.
[16]	Sung F, Yang Y, Zhang L, Xiang T, Torr P H, Hospedales T M. Learning to compare: Relation network for few-shot learning. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Salt Lake City, USA: IEEE, 2018. 1199−1208
[17]	Hao F, He F, Cheng J, Wang L, Cao J, Tao D. Collect and select: Semantic alignment metric learning for few-shot learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. Seoul, South Korea: IEEE, 2019. 8460−8469
[18]	Li W, Xu J, Huo J, Wang L, Gao Y, Luo J. Distribution consistency based covariance metric networks for few-shot learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. Honolulu, USA: 2019. 8642−8649
[19]	Zhu W, Li W, Liao H, Luo J. Temperature network for few-shot learning with distribution-aware large margin metric. Pattern Recognition, 2021, 112: Article No. 109381
[20]	Allen K, Shelhamer E, Shin H, Tennbaum J B. Infinite mixture prototypes for few-shot learning. In: Proceedings of the International Conference on Machine Learning. Long Beach, USA: ICML, 2019. 232−241
[21]	Liu J, Song L, Qin Y. Prototype rectification for few-shot learning. In: Proceedings of the European Conference Computer Vision. Glasgow, United Kingdom: Springer, 2020. 741−756
[22]	Li X, Li Y, Zheng Y, Zhu R, Ma Z, Xue J, et al. ReNAP: Relation network with adaptive prototypical learning for few-shot classification. Neurocomputing, 2023, 520: 356−364 doi: 10.1016/j.neucom.2022.11.082
[23]	Vinyals O, Blundell C, Lillicrap T, Kavukcuoglu K, Wierstra D. Matching networks for one shot learning. In: Proceedings of the 30th International Conference on Neural Information Processing Systems. Barcelona, Spain: 2016. 3637−3645
[24]	Bertinetto L, Henriques J, Torr P, Vedaldi A. Meta-learning with differentiable closed-form solvers. In: Proceedings of the International Conference on Learning Representations. New Orleans, USA: ICLR, 2019. 1−13
[25]	Oreshkin B, Rodríguez Loópez P, Lacoste A. TADAM: Task dependent adaptive metric for improved few-shot learning. In: Proceedings of the 32th International Conference on Neural Information Processing Systems. Montréal, Canada: 2018. 719−729
[26]	Russakovsky O, Deng J, Su H, Krause J, Satheesh S, Ma S, et al. Imagenet large scale visual recognition challenge. International Journal of Computer Vision, 2015, 115: 211−252 doi: 10.1007/s11263-015-0816-y
[27]	Ravi S, Larochelle H. Optimization as a model for few-shot learning. In: Proceedings of the International Conference on Learning Representations. San Juan, Puerto Rico: ICLR, 2016. 1−11
[28]	Krizhevsky A. Learning Multiple Layers of Features from Tiny Images, [Master thesis], Toronto University, Canda, 2009.
[29]	Bertinetto L, Henriques J F, Valmadre J, Torr P H S, Vedaldi A. Learning feed-forward one-shot learners. In: Proceedings of the 30th International Conference on Neural Information Processing Systems. Barcelona, Spain: 2016. 523−531
[30]	Qiao L, Shi Y, Li J, Wang Y, Huang T, Tian Y. Transductive episodic-wise adaptive metric for few-shot learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. Seoul, South Korea: IEEE, 2019. 3603−3612
[31]	Bateni P, Barber J, Van de Meent J W, Wood F. Enhancing few-shot image classification with unlabelled examples. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. Waikoloa, USA: IEEE, 2022. 2796−2805
[32]	Wang B, Li L, Verma M, Nakashima Y, Kawasaki R, Nagahara H. Match them up: visually explainable few-shot image classification. Applied Intelligence, 2023, 53: 10956−10977 doi: 10.1007/s10489-022-04072-4
[33]	Ji Z, Chai X, Yu Y, Pang Y, Zhang Z. Improved prototypical networks for few-shot learning. Pattern Recognition Letters, 2020, 140: 81−87 doi: 10.1016/j.patrec.2020.07.015
[34]	Shen X, Xiao Y, Hu S X, Sbai O, Aubry M. Re-ranking for image retrieval and transductive few-shot classification. In: Proceedings of the 35th International Conference on Neural Information Processing Systems. Virtual Event: 2021. 25932−25943
[35]	Gao F, Cai L, Yang Z, Song S, Wu C. Multi-distance metric network for few-shot learning. International Journal of Machine Learning and Cybernetics, 2022, 13(9): 2495−2506 doi: 10.1007/s13042-022-01539-1
[36]	Ravichandran A, Bhotika R, Soatto S. Few-shot learning with embedded class models and Shot-Free meta training. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. Seoul, South Korea: IEEE, 2019. 331−339
[37]	Zhang C, Cai Y, Lin G, Shen C. DeepEMD: Few-shot image classification with differentiable earth mover's distance and structured classifiers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Seattle, USA: IEEE, 2020. 12203−12213
[38]	Simon C, Koniusz P, Nock R, Harandi M. Adaptive subspaces for few-shot learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Seattle, USA: IEEE, 2020. 4136−4145
[39]	Wang H, Zhao H, Li B. Bridging multi-task learning and meta-learning: Towards efficient training and effective adaptation. In: Proceedings of the International Conference on Machine Learning. Vienna, Austria: ICML, 2021. 10991−11002
[40]	Yan L, Li F, Zhang L, Zheng X. Discriminant space metric network for few-shot image classification. Applied Intelligence, 2023, 53: 17444−17459 doi: 10.1007/s10489-022-04413-3
[41]	Wang Y, Chao W L, Weinberger K Q, Maaten L. SimpleShot: Revisiting nearest-neighbor classification for few-shot learning. arXiv preprint arXiv: 1911.04623, 2019.
[42]	Dhillon G S, Chaudhari P, Ravichandran A, Soattoet S. A baseline for few-shot image classification. In: Proceedings of the International Conference on Learning Representations. Addis Ababa, Ethiopia: ICLR, 2020. 1−12