特征增强的小样本持续学习方法

Feature-enhanced Method for Few-shot Continual Learning

摘要

持续学习旨在使模型在动态变化的环境中按顺序学习一系列任务,同时避免在学习新任务时遗忘先前获得的知识.然而,在图像分类任务中,当每个新任务仅有极其有限的标注样本时,即在小样本场景下,持续学习面临严峻的挑战:一方面,数据稀缺加剧了灾难性遗忘问题;另一方面,模型难以从少量样本中学习到鲁棒且泛化性强的特征表示.为应对上述挑战,本文提出了一种特征增强的小样本持续学习方法.该方法通过引入频域注意力机制优化特征图谱中关键频段的响应权重,使模型能够在频域空间实现细粒度信息提取,提升对判别性特征的聚焦能力和泛化性能;设计动态核权重分配机制,自适应调节卷积层感受野,增强模型对多尺度特征的捕获能力,减轻灾难性遗忘的影响;在特征空间中增加几何约束,保持新旧类别决策边界的可区分性,促进模型在学习新类别时对旧类别知识进行有效的整合,缓解特征混淆现象.在小样本图像分类数据集上与多种对比方法进行了对比实验,结果表明所提方法在所有数据集的每个增量阶段都获得了最好的准确率,为解决数据受限环境下的图像持续学习问题提供了一种切实可行的解决方案.

Abstract

Continual learning aims to enable models to sequentially learn a series of tasks in dynamically changing environments while avoiding the forgetting of previously acquired knowledge when learning new tasks.However,in image classification tasks,when each new task contains only a very limited number of labeled samples,i.e.,in few-shot scenarios,continual learning faces severe challenges:on the one hand,data scarcity exacerbates the problem of catastrophic forgetting;on the other hand,the model struggles to learn robust and generalizable feature representations from only a few samples.To address these challenges,this paper proposes a feature-enhanced few-shot continual learning method.Specifically,a frequency-domain attention mechanism is introduced to optimize the response weights of key frequency bands in the feature maps,enabling fine-grained information extraction in the frequency space and improving the model′s focus on discriminative features and its generalization ability;a dynamic kernel weight allocation mechanism is designed to adaptively adjust the receptive fields of convolutional layers,thereby enhancing the capture of multi-scale features and alleviating catastrophic forgetting;and additional geometric constraints are imposed in the feature space to maintain the separability of decision boundaries between old and new classes,which facilitates the effective integration of previous knowledge when learning new categories and mitigates feature confusion.Extensive experiments on the benchmark few-shot image classification datasets demonstrate that the proposed method consistently achieves the best accuracy at every incremental stage,providing a practical and effective solution to continual learning under data-constrained environments.

关键词

小样本持续学习 / 特征增强 / 深度神经网络 / 注意力机制 / 知识蒸馏 / 对比学习

Key words

few-shot continual learning / feature enhancement / deep neural networks / attention mechanism / knowledge distillation / contrastive learning

[1] Tan M X,Pang R M,Le Q V.Efficientdet:scalable and efficient object detection[C]//IEEE Conference on Computer Vision and Pattern Recognition,2020:10781-10790.
[2] Zhou D W,Ye H J,Zhan D C.Learning placeholders for open-set recognition[C]//IEEE Conference on Computer Vision and Pattern Recognition,2021:4401-4410.
[3] French R M.Catastrophic forgetting in connectionist networks[J].Trends in Cognitive Sciences,1999,3(4):128-135.
[4] David L P,Marc A R.Gradient episodic memory for continual learning[J].Advances in Neural Information Processing Systems,2017,30:6467-6476,doi:10.5555/3295222.3295393.
[5] Rebuffi S,Kolesnikov A,Sperl G,et al.icarl:incremental classifier and representation learning[C]//IEEE Conference on Computer Vision and Pattern Recognition,2017:2001-2010.
[6] Yang Y,Zhou D W,Zhang D C,et al.Adaptive deep models for incremental learning:considering capacity scalability and sustainability[C]//Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining,2019:74-82.
[7] Wang F Y,Zhou D W,Liu L,et al.Beef:bi-compatible class-incremental learning via energy-based expansion and fusion[C/OL]//11th International Conference on Learning Representations,https://api.semanticscholar.org/CorpusID:259298459,2023.
[8] Tao X Y,Hong X P,Chang X Y,et al.Few-shot class-incremental learning[C]//IEEE Conference on Computer Vision and Pattern Recognition,2020:12183-12192.
[9] Zhao H B,Fu Y J,Kang M T,et al.Mgsvf:multi-grained slow versus fast framework for few-shot class-incremental learning[J].IEEE Transactions on Pattern Analysis and Machine Intelligence,2021,46(3):1576-1588.
[10] Zhang C,Song N,Lin G S,et al.Few-shot incremental learning with continually evolved classifiers[C]//IEEE Conference on Computer Vision and Pattern Recognition,2021:12455-12464.
[11] Cui Y W,Deng W X,Xu X,et al.Uncertainty-guided semi-supervised few-shot class-incremental learning with knowledge distillation[J].IEEE Transactions on Multimedia,2022,25:6422-6435,doi:10.1109/TMM.2022.3208743.
[12] Fu Z L,Wang Z,Xu X L,et al.Knowledge aggregation networks for class incremental learning[J].Pattern Recognition,2023,137:109310,doi:10.1016/j.patcog.2023.109310.
[13] Duan J S,Zhang X.A class-incremental learning method for interactive event detection via interaction,contrast and distillation[J].Applied Sciences,2024,14(19):87-88.
[14] Li D P,Wang T Q,Chen J W,et al.Multi-view class incremental learning[J].Information Fusion,2024,102:102021,doi:10.1016/j.inffus.2023.102021.
[15] Mahmoodi L,Moghadam P,Hayat M,et al.Flashbacks to harmonize stability and plasticity in continual learning[J].Neural Networks,2025,190:107616,doi:10.1016/j.neunet.2025.107616.
[16] Gao H W,Huo X,Zhu C,et al.Task similarity-based continual learning for multi-phase environments and its application in few-shot fault diagnosis[J].Mechanical Systems and Signal Processing,2025,235:112862,doi:10.1016/j.ymssp.2025.112862.
[17] Xu R,Huang J T,Li Y Q,et al.FSCA:few-shot learning via embedding adaptation with corner multi-head attention[J].Electronics,2024,14(1):130,doi:10.3390/electronics.14010130.
[18] Tang Y H,Li G,Zhang M,et al.Few-shot learning based on dimensionally enhanced attention and logit standardization self-distillation[J].Electronics,2024,13(15):2928,doi:10.3390/electronics.13152928.
[19] Zhao P,Wang L,Zhao X Y,et al.Few-shot learning based on prototype rectification with a self-attention mechanism[J].Expert Systems with Applications,2024,249:123586,doi:10.1016/j.eswa.2024.123586.
[20] Zheng Z J,Zhu Y Y,Wu H,et al.SGE:semantic-guided generalization enhancement for few-shot learning[J].Knowledge-Based Systems,2025,323:113761,doi:10.1016/j.knosys.2025.113761.
[21] Liu Y K,Wei X J,Shi D M,et al.Towards few-shot learning with triplet metric learning and Kullback-Leibler optimization[J].Complex & Intelligent Systems,2025,11(7):313,doi:10.1007/s40747-025-01935-4.
[22] Li Y,Li G S.TDLPN:transductive dual label propagation network for few-shot learning[J].Engineering Letters,2025,33(6):1806-1814.
[23] Zhao J,Cao Z Y,Wang H L,et al.ProFusion:multimodal prototypical networks for few-shot learning with feature fusion[J].Symmetry,2025,17(5):796,doi:10.3390/sym.17050796.
[24] YAO G L,ZHU J T,ZHOU W L,et al.Few-shot class-incremental learning based on feature distribution learning[J].Computer Engineering and Applications,2023,59(14):151-157.
[25] Paeedeh N,Pratama M,Wibirama S,et al.Few-shot class incremental learning via robust transformer approach[J].Information Sciences,2024,675:120751,doi:10.1016/j.ins.2024.120751.
[26] Akmel F,Meng F M,Wu Q B,et al.Class similarity weighted knowledge distillation for few shot incremental learning[J].Neurocomputing,2024,584:127587,doi:10.1016/j.neucom.2024.127587.
[27] Fu Z L,Tang D F,Ma P C,et al.Distributed few-shot learning with prototype distribution correction[J].Applied Intelligence,2023,53(24):30552-30565.
[28] Akmel F D,Meng F M,Liu M Y,et al.Few-shot class incremental learning via prompt transfer and knowledge distillation[J].Image and Vision Computing,2024,151:105251,doi:10.1016/j.imavis.2024.105251.
[29] Wang X,Ji Z,Yu Y L,et al.Model attention expansion for few-shot class-incremental learning[J].IEEE Transactions on Image Processing,2024,33:4419-4431,doi:10.1109/TIP.2024.3434475.
[30] Wang X,Ji Z,Pang Y W,et al.A cognition-driven framework for few-shot class-incremental learning[J].Neurocomputing,2024,600:128118,doi:10.1016/j.neucom.2024.128118.
[31] Tu D W,Yi H Y,Zhang T Y,et al.Embedding space allocation with angle-norm joint classifiers for few-shot class-incremental learning[J].Neural Networks,2025,190:107608,doi:10.1016/j.neunet.2025.107608.
[32] Chen L W,Gu L,Zheng D Z,et al.Frequency-adaptive dilated convolution for semantic segmentation[C]//IEEE Conference on Computer Vision and Pattern Recognition,2024:3414-3425.
[33] Li X,Wang W H,Hu X L,et al.Selective kernel networks[C]//IEEE Conference on Computer Vision and Pattern Recognition,2019:510-519.
[34] Subramanian A,Sizikova E,Majaj N,et al.Spatial-frequency channels,shape bias,and adversarial robustness[J].Advances in Neural Information Processing Systems,2023,36:4137-4149,doi:10.5555/3666122.3666303.
[35] Weigend A S,Mangeas M,Srivastava A N.Nonlinear gated experts for time series:discovering regimes and avoiding overfitting[J].International Journal of Neural Systems,1995,6(4):373-399.
[36] Santos C F G D,Papa J P.Avoiding overfitting:a survey on regularization methods for convolutional neural networks[J].ACM Computing Surveys,2022,54(10s):1-25.
[37] Wen Y D,Zhang K P,Li Z F,et al.A discriminative feature learning approach for deep face recognition[C]//Computer Vision,14th European Conference,2016:499-515.
[38] Song O H,Xiang Y,Jegelka S,et al.Deep metric learning via lifted structured feature embedding[C]//IEEE Conference on Computer Vision and Pattern Recognition,2016:4004-4012.
[39] Vladimir N Vapnik.An overview of statistical learning theory[J].IEEE Transactions on Neural Networks,1999,10(5):988-999.
[40] Bartlett P L,Mendelson S.Rademacher and Gaussian complexities:risk bounds and structural results[J].Journal of Machine Learning Research,2002,3(11):463-482.
[41] Mohri M,Rostamizadeh A,Talwalkar A.Foundations of machine learning[M].MIT Press,2012.
[42] Koltchinskii V,Panchenko D.Empirical margin distributions and bounding the generalization error of combined classifiers[J].The Annals of Statistics,2002,30(1):1-50.
[43] Krizhevsky A.Learning multiple layers of features from tiny images[D].Toronto:University of Toronto,2009.
[44] Wah C,Branson S,Welinder P,et al.The caltech-ucsd birds-200-2011 dataset[OL].http://www.birdfieldguide.com,2011.
[45] Russakovsky O,Deng J,Su H,et al.Imagenet large scale visual recognition challenge[J].International Journal of Computer Vision,2015,115(3):211-252.
[46] Borgli H,Thambawita V,Smedsrud H P.HyperKvasir,a comprehensive multi-class image and video dataset for gastrointestinal endoscopy[J].Scientific Data,2020,7(1):283,doi:10.1038/841597-020-00622-y.
[47] Zhou D W,Wang F Y,Ye H J,et al.Forward compatible few-shot class-incremental learning[C]//IEEE Conference on Computer Vision and Pattern Recognition,2022:9046-9056.
[48] He K M,Zhang X Y,Ren S Q,et al.Deep residual learning for image recognition[C]//IEEE Conference on Computer Vision and Pattern Recognition,2016:770-778.
[49] Zhu F,Cheng Z,Zhang X Y,et al.Class-incremental learning via dual augmentation[C]//Advances in Neural Information Processing Systems,2021:14306-14318.
[50] Zhu F,Zhang X Y,Wang C,et al.Prototype augmentation and self-supervision for incremental learning[C]//IEEE/CVF Conference on Computer Vision and Pattern Recognition,2021:5871-5880.
[51] Wang Q W,Zhou D W,Zhang Y K,et al.Few-shot class-incremental learning via training-free prototype calibration[J].Advances in Neural Information Processing Systems,2023,36:15060-15076,doi:10.5555/3666122.3666783.
[52] Guo C X,Zhao Q,Lyu S C,et al.Decision boundary optimization for few-shot class-incremental learning[C]//IEEE/CVF International Conference on Computer Vision,2023:3501-3511.
[53] Deng Y,Xiang X.Expanding hyperspherical space for few-shot class-incremental learning[C]//IEEE/CVF Winter Conference on Applications of Computer Vision,2024:1967-1976.
[54] Wei Z H,Shi J X,Yang J,et al.VIP-FSCIL:a more robust approach for FSCIL[C]//IEEE International Conference on Multimedia and Expo,2024:1-6.
[55] Yang Y C,Jin L Y,Zhang S,et al.Separable and flexible classification on pseudo-features for few-shot class-incremental learning[C]//International Conference on Multimedia Retrieval,2025:1672-1681.
附中文参考文献:
[24] 姚光乐,祝钧桃,周文龙,等.基于特征分布学习的小样本类增量学习[J].计算机工程与应用,2023,59(14):151-157.

基金资助

国家自然科学基金项目(62206161,U21A20513,62476157,62406179)资助;山西省重点研发项目(202202020101003)资助;山西省基础研究计划项目(202403021222026)资助.