计算机科学

计算机科学 ›› 2026, Vol. 53 ›› Issue (9): 261-270.doi: 10.11896/jsjkx.260600084

• 计算机图形学&多媒体 • 上一篇    下一篇

基于动态协同的测试时自适应长尾识别框架

刘晓燕1, 程春怀2, 张云峰1, 包芳勋3   

  1. 1 山东财经大学计算机与人工智能学院 济南 250014
    2 山东省教育招生考试院 济南 250014
    3 山东大学数学学院 济南 250014
  • 收稿日期:2026-05-12 修回日期:2026-07-03 出版日期:2026-09-15 发布日期:2026-09-10
  • 通讯作者: 张云峰(yfzhang@sdufe.edu.cn)
  • 作者简介:(242215040@mail.sdufe.edu.cn)
  • 基金资助:
    山东省创新联合基金(ZR2024LZH00);国家自然科学基金(62506209)

Test-time Adaptive Framework with Dynamic Collaboration for Long-tailed Visual Recognition

LIU Xiaoyan1, CHENG Chunhuai2, ZHANG Yunfeng1, BAO Fangxun3   

  1. 1 School of Computer Science and Artificial Intelligence,Shandong University of Finance and Economics,Jinan 250014,China
    2 Shandong Provincial Education Admissions Examination Institute,Jinan 250014,China
    3 School of Mathematics,Shandong University,Jinan 250014,China
  • Received:2026-05-12 Revised:2026-07-03 Published:2026-09-15 Online:2026-09-10
  • About author:LIU Xiaoyan,born in 2000,postgra-duate.Her main research interests include long-tailed visual recognition,test-time adaptation and deep learning.
    ZHANG Yunfeng,born in 1977,Ph.D,professor,doctoral supervisor.His main research interests include computer vision,data analysis,image processing and visual analytics.
  • Supported by:
    Shandong Provincial Innovation Joint Fund(ZR2024LZH00) and National Natural Science Foundation of China(62506209).

PDF (PC)

253

摘要： 现实视觉识别任务在训练与测试阶段面临双重挑战:训练数据普遍呈现显著的长尾分布,导致模型容易偏向头部类别;而在测试无关(test-agnostic)场景中,测试阶段的类别分布通常未知且可能动态变化,使得依赖测试先验或静态聚合的传统方法难以保持稳定的性能。现有基于多专家的方法仍存在专家特化粒度粗、测试时自适应参数量大、尾部类别路由效率低等问题。针对上述问题,提出一种基于动态协同的测试时自适应长尾识别框架(Dynamic Collaborative Framework,DCF)。该方法采用共享骨干网络与微专家头相结合的结构,通过细粒度专家建模提升对不同类别区域的特化能力;同时,设计带先验校正的分组训练目标,并结合预测级多样性正则化与分类头正交性正则化,促进专家之间形成动态协同的互补表示。在测试阶段,进一步提出一种仅优化组级聚合权重的轻量级测试时自适应机制,无须更新骨干网络和专家参数即可实现对未知测试分布的动态适应;此外,还给出一种基于组原型的快速聚合变体,以兼顾推理效率。多个主流长尾识别基准数据集上的实验表明,所提方法在多数测试分布下达到或超过现有先进方法,尤其在尾部偏置场景下表现出更强的鲁棒性。消融实验进一步验证了专家协同设计、双重正则化和测试时自监督优化的有效性。

关键词: 长尾识别, 测试无关, 动态协同, 测试时自适应, 微专家

Abstract: Real-world visual recognition tasks face dual challenges during training and testing:training data usually exhibit a pronounced long-tailed distribution,which causes models to be biased toward head classes;meanwhile,in test-agnostic scenarios,the class distribution at test time is typically unknown and may change dynamically,making it difficult for traditional methods that rely on test priors or static aggregation to maintain stable performance.Existing multi-expert methods still suffer from coarse-grained expert specialization,excessive test-time adaptation parameters,and low routing efficiency for tail classes.To address these issues,this paper proposes a test-time adaptive framework with dynamic collaboration for long-tailed visual recognition,termed Dynamic Collaborative Framework(DCF).The proposed method adopts a shared backbone network with multiple micro-expert heads to improve specialization across different class regions through fine-grained expert modeling;meanwhile,it designs a group-wise training objective with prior correction and incorporates prediction-level diversity regularization as well as classifier-head orthogonality regularization to promote dynamically collaborative complementary representations among experts.At test time,it further introduces a lightweight test-time adaptation mechanism that optimizes only group-level aggregation weights,enabling dynamic adaptation to unknown test distributions without updating the backbone network or expert parameters;in addition,a fast aggregation variant based on group prototypes is developed to balance inference efficiency.Experiments on multiple mainstream long-tailed recognition benchmark datasets show that the proposed method achieves or surpasses existing state-of-the-art methods under most test distributions and exhibits stronger robustness,particularly in tail-biased scenarios.Ablation studies further verify the effectiveness of the expert collaboration design,dual regularization,and test-time self-supervised optimization.

Key words: Long-tailed recognition, Test-agnostic, Dynamic collaboration, Test-time adaptation, Micro-experts

中图分类号: 

  • TP391
[1] 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.
[2] ZHOU B,LAPEDRIZA A,KHOSLA A,et al.Places:A 10 million image database for scene recognition[J].IEEE Transactions on Pattern Analysis and Machine Intelligence,2017,40(6):1452-1464.
[3] WEI X C.Progress and approaches in long-tailed image recognition[J].China Basic Science,2023(5):50-57.
[4] GAO J,CHEN J,FU H,et al.Dynamic mixup for multi-label long-tailed food ingredient recognition[J].IEEE Transactions on Multimedia,2022,25:4764-4773.
[5] LI M,HU Z,LU Y,et al.Feature fusion from head to tail for long-tailed visual recognition[C] //Proceedings of the AAAI Conference on Artificial Intelligence.Palo Alto:AAAI Press,2024:13581-13589.
[6] ZHANG M L,ZHANG X Y,WANG C,et al.Towards prior gap and representation gap for long-tailed recognition[J].Pattern Recognition,2023,133:109012.
[7] REN J,YU C,SHENG S,et al.Balanced meta-softmax for long-tailed visual recognition[C] //Advances in Neural Information Processing Systems.Red Hook:Curran Associates,2020,33:4175-4186.
[8] CAO K,WEI C,GAIDON A,et al.Learning imbalanced data-sets with label-distribution-aware margin loss[C] //Advances in Neural Information Processing Systems.Red Hook:Curran Associates,2019:1567-1578.
[9] MENON A K,JAYASUMANA S,RAWAT A S,et al.Long-tail learning via logit adjustment[PP/OL].arXiv(2020-07-14)[2026-08-06].https://arxiv.org/abs/2007.07314.
[10] ZHANG S,LI Z,YAN S,et al.Distribution alignment:A unified framework for long-tail visual recognition[C] //Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.Los Alamitos:IEEE Computer Society Press,2021:2361-2370.
[11] KANG B,LI Y,XIE S,et al.Exploring balanced feature spaces for representation learning[C] //Proceedings of the International Conference on Learning Representations.San Diego:Open-Review.net,2020.
[12] KANG B,XIE S,ROHRBACH M,et al.Decoupling representation and classifier for long-tailed recognition[PP/OL].arXiv(2019-10-21)[2026-08-06].https://arxiv.org/abs/1910.09217.
[13] GONG X,LV J R.Survey of test-time adaptation:From methods to applications[J].Journal of Sichuan Normal University(Natural Science),2025,48(6):711-722.
[14] ZHANG Y,HOOI B,HONG L,et al.Self-supervised aggregation of diverse experts for test-agnostic long-tailed recognition[C] //Advances in Neural Information Processing Systems.Red Hook:Curran Associates,2022:34077-34090.
[15] GUO C,CHEN W,HUANG A,et al.Prototype alignment with dedicated experts for test-agnostic long-tailed recognition[J].IEEE Transactions on Multimedia,2025,27:455-465.
[16] ZHOU L,LI N,YE M,et al.Source-free domain adaptation with class prototype discovery[J].Pattern Recognition,2024,145:109974.
[17] YANG X,CHEN X,LI M,et al.A versatile framework for continual test-time domain adaptation:Balancing discriminability and generalizability[C] //Proceedings of the IEEE/CVF Confe-rence on Computer Vision and Pattern Recognition.Los Alamitos:IEEE Computer Society Press,2024:23731-23740.
[18] SHAZEER N,MIRHOSEINI A,MAZIARZ K,et al.Outra-geously large neural networks:The sparsely-gated mixture-of-experts layer[PP/OL].arXiv(2017-01-23)[2026-08-06].https://arxiv.org/abs/1701.06538.
[19] LEPIKHIN D,LEE H,XU Y,et al.GShard:Scaling giant models with conditional computation and automatic sharding[PP/OL].arXiv(2020-06-30)[2026-08-06].https://arxiv.org/abs/2006.16668.
[20] CAI J,WANG Y,HWANG J N.ACE:Ally complementary experts for solving long-tailed recognition in one-shot[C] //Proceedings of the IEEE/CVF International Conference on Computer Vision.Los Alamitos:IEEE Computer Society Press,2021:112-121.
[21] WANG X,LIAN L,MIAO Z,et al.Long-tailed recognition by routing diverse distribution-aware experts[PP/OL].arXiv(2020-10-05)[2026-08-06].https://arxiv.org/abs/2010.01809.
[22] HONG Y,HAN S,CHOI K,et al.Disentangling label distribution for long-tailed visual recognition[C] //Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.Los Alamitos:IEEE Computer Society Press,2021:6626-6636.
[23] ZHAO Q,DAI Y,LI H,et al.LTGC:Long-tail recognition via leveraging LLMs-driven generated content[C] //Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.Los Alamitos:IEEE Computer Society Press,2024:19510-19520.
[24] ZHAO Q,DAI Y,LIN S,et al.LTRL:Boosting long-tail recognition via reflective learning[C] //Computer Vision-ECCV 2024.Cham:Springer,2024:1-18.
[25] ZHAO S,WEN X,LIU J,et al.Learning from neighbors:Category extrapolation for long-tail learning[C] //Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.Los Alamitos:IEEE Computer Society Press,2025:30483-30492.
[26] LIU Y,YANG L,WANG Y.Long-tailed classification withmulti-granularity semantics[C] //Proceedings of the IEEE/CVF International Conference on Computer Vision.Los Alamitos:IEEE Computer Society Press,2025:4285-4294.
[27] HINTON G,VINYALS O,DEAN J.Distilling the knowledge in a neural network[PP/OL].arXiv(2015-03-09)[2026-08-06].https://arxiv.org/abs/1503.02531.
[28] HEO B,KIM J,YUN S,et al.A comprehensive overhaul of feature distillation[C] //Proceedings of the IEEE/CVF International Conference on Computer Vision.Los Alamitos:IEEE Computer Society Press,2019:1921-1930.
[29] ZHAO B,CUI Q,SONG R,et al.Decoupled knowledge distillation[C] //Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.Los Alamitos:IEEE Computer Society Press,2022:11953-11962.
[30] ZHAO W,ZHU X,GUO K,et al.Grouped knowledge distillation for deep face recognition[C] //Proceedings of the AAAI Conference on Artificial Intelligence.Palo Alto:AAAI Press,2023:3615-3623.
[31] LI J,ZHOU P,XIONG C,et al.Prototypical contrastive learning of unsupervised representations[PP/OL].arXiv(2020-05-11)[2026-08-06].https://arxiv.org/abs/2005.04966.
[32] LI J,XIONG C,HOI S C H.MoPro:Webly supervised learning with momentum prototypes[PP/OL].arXiv(2020-09-17)[2026-08-06].https://arxiv.org/abs/2009.07995.
[33] YANG H M,ZHANG X Y,YIN F,et al.Convolutional prototype network for open set recognition[J].IEEE Transactions on Pattern Analysis and Machine Intelligence,2022,44(5):2358-2370.
[34] KRIZHEVSKY A.Learning multiple layers of features from tiny images[D].Toronto:University of Toronto,2009.
[35] LIU Z,MIAO Z,ZHAN X,et al.Large-scale long-tailed recogni-tion in an open world[C] //Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.Los Ala-mitos:IEEE Computer Society Press,2019:2537-2546.
[36] VAN HORN G,MAC AODHA O,SONG Y,et al.The iNaturalist species classification and detection dataset[C] //Procee-dings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.Los Alamitos:IEEE Computer Society Press,2018:8769-8778.
[37] ZHU J,WANG Z,CHEN J,et al.Balanced contrastive learning for long-tailed visual recognition[C] //Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.Los Alamitos:IEEE Computer Society Press,2022:6908-6917.
[38] ZHONG Z,CUI J,LIU S,et al.Improving calibration for long-tailed recognition[C] //Proceedings of the IEEE/CVF Confe-rence on Computer Vision and Pattern Recognition.Los Alamitos:IEEE Computer Society Press,2021:16489-16498.
[39] LI T,CAO P,YUAN Y,et al.Targeted supervised contrastive learning for long-tailed recognition[C] //Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.Los Alamitos:IEEE Computer Society Press,2022:6918-6928.
[40] HOU C,ZHANG J,WANG H,et al.Subclass-balancing cont-rastive learning for long-tailed recognition[C] //Proceedings of the IEEE/CVF International Conference on Computer Vision.Los Alamitos:IEEE Computer Society Press,2023:5372-5384.
[41] ALEXANDRIDIS K,LUO S,NGUYEN A,et al.Inverse image frequency for long-tailed image recognition[J].IEEE Transactions on Image Processing,2023,32:5721-5736.
Viewed
Full text


Abstract

Cited
  Shared   
  Discussed   
No Suggested Reading articles found!
渝ICP备16013121号-4
地址：重庆市两江新区洪湖西路18号    邮编：401121  
电话：023-63500828（编辑部） 023-67039612（会议/专题） 023-67039623（财务/发票）
E-mail：jsjkx12@163.com
本系统由北京玛格泰克科技发展有限公司设计开发