Agent prototype distillation algorithm for few-shot object detection

doi:10.11772/j.issn.1001-9081.2025020142

Journal of Computer Applications ›› 2026, Vol. 46 ›› Issue (1): 233-241.DOI: 10.11772/j.issn.1001-9081.2025020142

• Multimedia computing and computer simulation • Previous Articles Next Articles

Agent prototype distillation algorithm for few-shot object detection

Binhong XIE, Rui WANG(), Rui ZHANG, Yingjun ZHANG

College of Computer Science and Technology，Taiyuan University of Science and Technology，Taiyuan Shanxi 030024，China

Received:2025-02-14 Revised:2025-03-18 Accepted:2025-03-24 Online:2026-01-10 Published:2026-01-10
Contact: Rui WANG
About author:XIE Binhong， born in 1971， M. S.， professor. His research interests include intelligent software engineering， machine learning.
ZHANG Rui， born in 1987， Ph. D.， professor. His research interests include intelligent information processing.
ZHANG Yingjun， born in 1969， M. S.， professor-level senior engineer. His research interests include intelligent perception and decision-making.
Supported by:
Fundamental Research Program of Shanxi Province(20210302123216);Key Research and Development Program for the Introduction of High-Level Scientific and Technological Talents in Lvliang City(2022RC08);Shanxi Province Industry-Education Integration Postgraduate Joint Training Demonstration Base Project(2022JD11)

代理原型蒸馏的小样本目标检测算法

谢斌红, 王瑞(), 张睿, 张英俊

太原科技大学计算机科学与技术学院，太原 030024

通讯作者: 王瑞
作者简介:谢斌红（1971—），男，山西太原人，教授，硕士，CCF会员，主要研究方向：智能化软件工程、机器学习
张睿（1987—），男，山西太原人，教授，博士，主要研究方向：智能信息处理
张英俊（1969—），男，山西太原人，教授级高级工程师，硕士，主要研究方向：智能感知决策。
基金资助:
山西省基础研究计划项目(20210302123216);吕梁市引进高层次科技人才重点研发项目(2022RC08);山西省产教融合研究生联合培养示范基地项目(2022JD11)

Abstract

Abstract:

The existing Few-Shot Object Detection （FSOD） algorithms are constrained by insufficient class-level prototype generation accuracy and loss of detail information， which limit the feature representation capability in target regions. To address this issue， an Agent Prototype Aggregation （APA） based FSOD algorithm named APA-FSOD was proposed. In the algorithm， the support features were distilled into detail-rich prototypes through agent attention， and the prototype vectors were assigned to the query feature map accurately based on their correlations， thereby enhancing the feature representation capability in target instance regions significantly. Additionally， a Wavelet Convolution Enhancement Module （WCEM） and an Adaptive Multi-Relation Fusion （AMRF） module were designed to optimize global feature extraction and advanced feature fusion of the algorithm， respectively. Experimental results demonstrate that on three novel class splits of the PASCAL VOC dataset， the nAP50 of APA-FSOD is improved by 0.5 to 1.1 percentage points compared to that of the baseline method VFA （Variational Feature Aggregation）； under the 30-shot setting of the MS COCO dataset， nAP of APA-FSOD is increased by 1.0 percentage point compared to that of the meta-learning method SMPCCNet （Support-query Mutual Promotion and Classification Correction Network）. It can be seen that the proposed algorithm achieves significant accuracy improvement in FSOD.

Key words: Few-Shot Object Detection (FSOD), meta-learning, agent prototype distillation, wavelet convolution, multi-relation fusion

摘要：

针对现有小样本目标检测（FSOD）算法中类级原型生成精度不足以及细节信息缺失导致的目标区域特征表达能力受限的问题，提出一种基于代理原型聚合（APA）的FSOD算法APA-FSOD。该算法通过代理注意力将支持特征蒸馏为细节丰富的原型，并基于原型向量的相关性实现原型向量在查询特征图上的精准分配，从而显著强化目标实例区域的特征表达能力。此外，设计小波卷积增强模块（WCEM）和自适应多关系融合模块（AMRF），并分别用于优化算法的全局特征提取和高级特征融合。实验结果表明，在PASCAL VOC数据集的3种新类划分下， APA-FSOD的nAP50相较于基线方法VFA （Variational Feature Aggregation）提升了0.5~1.1个百分点；而在MS COCO数据集的30-shot设置下，与元学习方法SMPCCNet （Support-query Mutual Promotion and Classification Correction Network）相比， nAP提升了1.0个百分点。可见，所提算法显著提高了FSOD的精度。

关键词: 小样本目标检测, 元学习, 代理原型蒸馏, 小波卷积, 多关系融合

CLC Number:

TP399

Binhong XIE, Rui WANG, Rui ZHANG, Yingjun ZHANG. Agent prototype distillation algorithm for few-shot object detection[J]. Journal of Computer Applications, 2026, 46(1): 233-241.

谢斌红, 王瑞, 张睿, 张英俊. 代理原型蒸馏的小样本目标检测算法[J]. 《计算机应用》唯一官方网站, 2026, 46(1): 233-241.

Figures/Tables 13

Fig. 1 Training flow of few-shot object detection

Fig. 2 Detection flow of APA-FSOD

Fig. 3 Agent prototype aggregation module

Fig. 4 Wavelet convolution enhancement module

Tab. 1 Comparison of model performance （nAP50） on VOC 2007 test set under three novel class splits

范式	方法	Novel Set 1					Novel Set 2					Novel Set 3
范式	方法	$K n$ =1	$K n$ =2	$K n$ =3	$K n$ =5	$K n$ =10	$K n$ =1	$K n$ =2	$K n$ =3	$K n$ =5	$K n$ =10	$K n$ =1	$K n$ =2	$K n$ =3	$K n$ =5	$K n$ =10
Fine-tuning	TFA w/cos^［18］	39.8	36.1	44.7	55.7	56.0	23.5	26.9	34.1	35.1	39.1	30.8	34.8	42.8	49.5	49.8
	MPSR^［13］	41.7	—	51.4	55.2	61.8	24.4	—	39.2	39.9	47.8	35.6	—	42.3	48.0	49.7
	Retent RCNN^［21］	42.4	45.8	45.9	53.7	56.1	21.7	27.8	35.2	37.0	40.3	30.2	37.6	43.0	49.7	50.1
	SRR-FSD^［22］	47.8	50.5	51.3	55.2	56.8	32.5	35.3	39.1	40.8	43.8	40.1	41.5	44.3	46.9	46.4
	DeFRCN^［23］	53.6	57.5	61.5	64.1	60.8	30.1	38.1	47.0	53.3	47.9	48.4	50.9	52.3	54.9	57.4
	LVC^［24］	54.5	53.2	58.8	63.2	65.7	32.8	29.2	50.7	49.8	50.6	48.4	52.7	55.0	59.6	59.6
	CFA-DeFRCN^［20］	58.2	63.3	65.8	68.9	67.1	37.1	45.5	51.3	55.2	53.8	54.7	57.8	56.9	60.0	63.3
Meta-learning	DCNet^［16］	33.9	37.4	43.7	51.1	59.6	23.2	24.8	30.6	36.7	46.6	32.3	34.9	39.7	42.6	50.7
	CAReD^［25］	36.5	45.2	47.1	50.8	58.8	26.4	31.0	37.9	43.5	51.1	20.2	33.8	41.6	48.3	55.3
	Meta DETR^［7］	40.6	51.4	58.0	59.2	63.6	37.0	36.6	43.7	49.1	54.6	41.6	45.9	52.7	58.9	60.6
	QA-FewDet^［26］	42.4	51.9	55.7	62.6	63.4	25.9	37.8	46.6	48.9	51.1	35.2	42.9	47.8	54.8	53.5
	Meta FR-CNN^［15］	43.0	54.5	60.6	66.1	65.4	27.7	35.5	46.1	47.8	51.4	40.6	46.4	53.4	59.9	58.6
	SMPCCNet^［27］	45.5	50.3	54.2	56.1	—	27.9	35.6	43.7	46.2	—	37.7	44.5	48.2	55.6	—
	FM-FSOD^［28］	47.5	50.6	53.4	57.2	59.9	37.9	39.8	43.6	46.7	49.8	33.4	36.3	40.5	46.8	48.9
	FPD^［8］	48.1	62.2	64.0	67.6	68.4	29.8	43.2	47.7	52.0	53.9	44.9	53.8	58.1	61.6	62.9
	VFA^［3］	57.7	64.6	64.7	67.2	67.4	41.4	46.2	51.1	51.8	51.6	48.9	54.8	56.6	59.0	58.9
	APA-FSOD	58.2	65.1	65.4	68.0	68.1	42.0	46.7	51.8	52.7	52.4	49.4	55.4	57.5	60.1	59.7

Tab. 1 Comparison of model performance （nAP50） on VOC 2007 test set under three novel class splits

范式	方法	Novel Set 1					Novel Set 2					Novel Set 3
范式	方法	$K n$ =1	$K n$ =2	$K n$ =3	$K n$ =5	$K n$ =10	$K n$ =1	$K n$ =2	$K n$ =3	$K n$ =5	$K n$ =10	$K n$ =1	$K n$ =2	$K n$ =3	$K n$ =5	$K n$ =10
Fine-tuning	TFA w/cos^［18］	39.8	36.1	44.7	55.7	56.0	23.5	26.9	34.1	35.1	39.1	30.8	34.8	42.8	49.5	49.8
	MPSR^［13］	41.7	—	51.4	55.2	61.8	24.4	—	39.2	39.9	47.8	35.6	—	42.3	48.0	49.7
	Retent RCNN^［21］	42.4	45.8	45.9	53.7	56.1	21.7	27.8	35.2	37.0	40.3	30.2	37.6	43.0	49.7	50.1
	SRR-FSD^［22］	47.8	50.5	51.3	55.2	56.8	32.5	35.3	39.1	40.8	43.8	40.1	41.5	44.3	46.9	46.4
	DeFRCN^［23］	53.6	57.5	61.5	64.1	60.8	30.1	38.1	47.0	53.3	47.9	48.4	50.9	52.3	54.9	57.4
	LVC^［24］	54.5	53.2	58.8	63.2	65.7	32.8	29.2	50.7	49.8	50.6	48.4	52.7	55.0	59.6	59.6
	CFA-DeFRCN^［20］	58.2	63.3	65.8	68.9	67.1	37.1	45.5	51.3	55.2	53.8	54.7	57.8	56.9	60.0	63.3
Meta-learning	DCNet^［16］	33.9	37.4	43.7	51.1	59.6	23.2	24.8	30.6	36.7	46.6	32.3	34.9	39.7	42.6	50.7
	CAReD^［25］	36.5	45.2	47.1	50.8	58.8	26.4	31.0	37.9	43.5	51.1	20.2	33.8	41.6	48.3	55.3
	Meta DETR^［7］	40.6	51.4	58.0	59.2	63.6	37.0	36.6	43.7	49.1	54.6	41.6	45.9	52.7	58.9	60.6
	QA-FewDet^［26］	42.4	51.9	55.7	62.6	63.4	25.9	37.8	46.6	48.9	51.1	35.2	42.9	47.8	54.8	53.5
	Meta FR-CNN^［15］	43.0	54.5	60.6	66.1	65.4	27.7	35.5	46.1	47.8	51.4	40.6	46.4	53.4	59.9	58.6
	SMPCCNet^［27］	45.5	50.3	54.2	56.1	—	27.9	35.6	43.7	46.2	—	37.7	44.5	48.2	55.6	—
	FM-FSOD^［28］	47.5	50.6	53.4	57.2	59.9	37.9	39.8	43.6	46.7	49.8	33.4	36.3	40.5	46.8	48.9
	FPD^［8］	48.1	62.2	64.0	67.6	68.4	29.8	43.2	47.7	52.0	53.9	44.9	53.8	58.1	61.6	62.9
	VFA^［3］	57.7	64.6	64.7	67.2	67.4	41.4	46.2	51.1	51.8	51.6	48.9	54.8	56.6	59.0	58.9
	APA-FSOD	58.2	65.1	65.4	68.0	68.1	42.0	46.7	51.8	52.7	52.4	49.4	55.4	57.5	60.1	59.7

Tab. 2 Comparison of base class accuracy under 1-shot setting

类别	FPD	VFA	APA-FSOD
平均	67.7	71.6	71.9
aeroplane	77.2	82.1	82.3
bicycle	77.0	81.7	81.8
boat	65.0	68.1	68.3
bottle	58.8	66.0	65.8
car	79.7	79.5	80.3
cat	84.0	86.6	87.3
chair	30.6	37.7	38.2
diningtable	46.6	55.0	55.7
dog	81.4	84.4	85.2
horse	84.3	86.7	86.8
person	76.3	77.3	77.6
pottedplant	43.9	45.9	46.0
sheep	72.8	75.4	75.6
train	74.2	80.9	80.5
tvmonitor	63.8	67.3	67.4

Tab. 3 Performance （nAP） comparison on COCO dataset

范式	方法	10-shot	30-shot
Fine-tuning	MPSR^［13］	9.8	14.1
	TFA w/cos^［18］	10.0	13.7
	Retent RCNN^［21］	10.5	13.8
	SRR-FSD^［22］	11.3	14.7
	TeSNet^［29］	14.1	16.5
Meta-learning	QA-FewDet^［26］	11.6	16.5
	Meta FR-CNN^［15］	12.7	16.6
	DCNet^［16］	12.8	18.6
	FM-FSOD^［28］	13.1	—
	CAReD^［25］	15.5	18.4
	VFA^［3］	16.2	18.9
	FPD^［8］	16.5	20.1
	SMPCCNet^［27］	16.8	18.5
	APA-FSOD	16.6	19.5

Tab. 4 Performance （nAP） comparison of different modules in ablation experiments at 3/5/10-shot under the novel set 3 of VOC dataset

WCEM	APA	AMRF	3-shot	5-shot	10-shot
			56.6	59.0	58.9
√			56.9	59.2	59.2
√	√		57.2	59.7	59.6
√	√	√	57.5	60.1	59.7

Tab. 5 Comparison of ablation experiment results for calculation costs

方法	参数量/10⁶	帧率/（frame·s^-1）	浮点计算量/GFLOPs
Baseline	56.18	14.6	843.63
GA	80.23	10.2	1 176.38
AA	81.74	12.4	965.16

Fig. 5 Impact of parameters N and M on few-shot detection performance

Fig. 6 Visualization of some detection results

Fig. 7 Attention heatmaps of feature queries

Fig. 8 Feature maps of query images

References 29

[1]	KÖHLER M， EISENBACH M， GROSS H M. Few-shot object detection： a comprehensive survey ［J］. IEEE Transactions on Neural Networks and Learning Systems， 2024， 35（9）： 11958-11978.
[2]	史燕燕，史殿习，乔子腾，等.小样本目标检测研究综述［J］.计算机学报， 2023， 46（8）： 1753-1780.
	SHI Y Y， SHI D X， QIAO Z T， et al. A survey on recent advances in few-shot object detection ［J］. Chinese Journal of Computers， 2023， 46（8）： 1753-1780.
[3]	HAN J， REN Y， DING J， et al. Few-shot object detection via variational feature aggregation ［C］// Proceedings of the 37th AAAI Conference on Artificial Intelligence. Palo Alto： AAAI Press， 2023： 755-763.
[4]	KANG B， LIU Z， WANG X， et al. Few-shot object detection via feature reweighting ［C］// Proceedings of the 2019 IEEE/CVF International Conference on Computer Vision. Piscataway： IEEE， 2019： 8419-8428.
[5]	YAN X， CHEN Z， XU A， et al. Meta R-CNN： towards general solver for instance-level low-shot learning ［C］// Proceedings of the 2019 IEEE/CVF International Conference on Computer Vision. Piscataway： IEEE， 2019： 9576-9585.
[6]	XIAO Y， LEPETIT V， MARLET R. Few-shot object detection and viewpoint estimation for objects in the wild ［J］. IEEE Transactions on Pattern Analysis and Machine Intelligence， 2023， 45（3）： 3090-3106.
[7]	ZHANG G， LUO Z， CUI K， et al. Meta-DETR： image-level few-shot detection with inter-class correlation exploitation ［J］. IEEE Transactions on Pattern Analysis and Machine Intelligence， 2023， 45（11）： 12832-12843.
[8]	WANG Z， YANG B， YUE H， et al. Fine-grained prototypes distillation for few-shot object detection ［C］// Proceedings of the 38th AAAI Conference on Artificial Intelligence. Palo Alto： AAAI Press， 2024： 5859-5866.
[9]	LU X， DIAO W， MAO Y， et al. Breaking immutable： information-coupled prototype elaboration for few-shot object detection ［C］// Proceedings of the 37th AAAI Conference on Artificial Intelligence. Palo Alto： AAAI Press， 2023： 1844-1852.
[10]	HAN D， YE T， HAN Y， et al. Agent Attention： on the integration of softmax and linear attention ［C］// Proceedings of the 2024 European Conference on Computer Vision， LNCS 15108. Cham： Springer， 2025： 124-140.
[11]	FINDER S E， AMOYAL R， TREISTER E， et al. Wavelet convolutions for large receptive fields ［C］// Proceedings of the 2024 European Conference on Computer Vision， LNCS 15112. Cham： Springer， 2025： 363-380.
[12]	李鸿天，史鑫昊，潘卫国，等.融合多尺度和注意力机制的小样本目标检测［J］.计算机应用， 2024， 44（5）： 1437-1444.
	LI H T， SHI X H， PAN W G， et al. Few-shot object detection via fusing multi-scale and attention mechanism ［J］. Journal of Computer Applications， 2024， 44（5）： 1437-1444.
[13]	WU J， LIU S， HUANG D， et al. Multi-scale positive sample refinement for few-shot object detection ［C］// Proceedings of the 2020 European Conference on Computer Vision， LNCS 12361. Cham： Springer， 2020： 456-472.
[14]	KARLINSKY L， SHTOK J， HARARY S， et al. RepMet： representative-based metric learning for classification and few-shot object detection ［C］// Proceedings of the 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway： IEEE， 2019： 5192-5201.
[15]	HAN G， HUANG S， MA J， et al. Meta Faster R-CNN： towards accurate few-shot object detection with attentive feature alignment ［C］// Proceedings of the 36th AAAI Conference on Artificial Intelligence. Palo Alto： AAAI Press， 2022： 780-789.
[16]	HU H， BAI S， LI A， et al. Dense relation distillation with context-aware aggregation for few-shot object detection ［C］// Proceedings of the 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway： IEEE， 2021： 10180-10189.
[17]	KINGMA D P， WELLING M. Auto-encoding variational Bayes ［EB/OL］. ［2024-12-20］. .
[18]	WANG X， HUANG T E， DARRELL T， et al. Frustratingly simple few-shot object detection ［C］// Proceedings of the 37th International Conference on Machine Learning. New York： JMLR.org， 2020： 9919-9928.
[19]	OpenMMLab. MMFewShot： OpenMMLab few shot learning toolbox and benchmark ［EB/OL］. ［2024-12-20］. .
[20]	GUIRGUIS K， HENDAWY A， ESKANDAR G， et al. CFA： constraint-based finetuning approach for generalized few-shot object detection ［C］// Proceedings of the 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway： IEEE， 2022： 4038-4048.
[21]	FAN Z， MA Y， LI Z， et al. Generalized few-shot object detection without forgetting ［C］// Proceedings of the 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway： IEEE， 2021： 4525-4534.
[22]	ZHU C， CHEN F， AHMED U， et al. Semantic relation reasoning for shot-stable few-shot object detection ［C］// Proceedings of the 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway： IEEE， 2021： 8778-8787.
[23]	QIAO L， ZHAO Y， LI Z， et al. DeFRCN： decoupled faster R-CNN for few-shot object detection ［C］// Proceedings of the 2021 IEEE/CVF International Conference on Computer Vision. Piscataway： IEEE， 2021： 8661-8670.
[24]	KAUL P， XIE W， ZISSERMAN A. Label， verify， correct： a simple few shot object detection method ［C］// Proceedings of the 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway： IEEE， 2022： 14217-14227.
[25]	QUAN J， GE B， CHEN L. Cross attention redistribution with contrastive learning for few shot object detection ［J］. Displays， 2022， 72： No.102162.
[26]	HAN G， HE Y， HUANG S， et al. Query adaptive few-shot object detection with heterogeneous graph convolutional networks ［C］// Proceedings of the 2021 IEEE/CVF International Conference on Computer Vision. Piscataway： IEEE， 2021： 3243-3252.
[27]	XU J， WANG Y， HE X， et al. Support-query mutual promotion and classification correction network for few-shot object detection ［J］. IEEE Signal Processing Letters， 2024， 31： 201-205.
[28]	LI Y， FENG W， LYU S， et al. Feature reconstruction and metric based network for few-shot object detection ［J］. Computer Vision and Image Understanding， 2023， 227： No.103600.
[29]	ZHAO X， LIU X， MA Y， et al. Temporal speciation network for few-shot object detection ［J］. IEEE Transactions on Multimedia， 2023， 25： 8267-8278.

Agent prototype distillation algorithm for few-shot object detection

代理原型蒸馏的小样本目标检测算法

RichHTML

PDF

Knowledge

Abstract

Cite this article

share this article

Figures/Tables 13

References 29

Related Articles 13

Recommended Articles

Metrics

[1]	Qiaoling QI, Xiaoxiao WANG, Qianqian ZHANG, Peng WANG, Yongfeng DONG. Label noise adaptive learning algorithm based on meta-learning [J]. Journal of Computer Applications, 2025, 45(7): 2113-2122.
[2]	Kun FU, Shicong YING, Tingting ZHENG, Jiajie QU, Jingyuan CUI, Jianwei LI. Graph data augmentation method for few-shot node classification [J]. Journal of Computer Applications, 2025, 45(2): 392-402.
[3]	Yanpeng ZHANG, Yuqian ZHAO, Fan ZHANG, Tenghai QIU, Gui GUI, Lingli YU. Capacitated vehicle routing problem solving method based on improved MAML and GVAE [J]. Journal of Computer Applications, 2025, 45(11): 3642-3648.
[4]	Zhihao WU, Ziqiu CHI, Ting XIAO, Zhe WANG. Meta-learning adaption for few-shot text-to-speech [J]. Journal of Computer Applications, 2024, 44(5): 1629-1635.
[5]	Keyi FU, Gaocai WANG, Man WU. Few-shot object detection method based on improved region proposal network and feature aggregation [J]. Journal of Computer Applications, 2024, 44(12): 3790-3797.
[6]	Xiaomin ZHOU, Fei TENG, Yi ZHANG. Automatic international classification of diseases coding model based on meta-network [J]. Journal of Computer Applications, 2023, 43(9): 2721-2726.
[7]	Hui WANG, Jianhong LI. Few-shot recognition method of 3D models based on Transformer [J]. Journal of Computer Applications, 2023, 43(6): 1750-1758.
[8]	Yaru HAN, Lianshan YAN, Tao YAO. Deep hashing retrieval algorithm based on meta-learning [J]. Journal of Computer Applications, 2022, 42(7): 2015-2021.
[9]	Yan LI, Jie GUO, Bin FAN. Feature construction and preliminary analysis of uncertainty for meta-learning [J]. Journal of Computer Applications, 2022, 42(2): 343-348.
[10]	WEI Chunwu, ZHAO Juanjuan, TANG Xiaoxian, QIANG Yan. Knowledge extraction method for follow-up data based on multi-term distillation network [J]. Journal of Computer Applications, 2021, 41(10): 2871-2878.
[11]	SUN Zhongfan, ZHOU Zhenghua, ZHAO Jianwei. Super-resolution reconstruction method with arbitrary magnification based on spatial meta-learning [J]. Journal of Computer Applications, 2020, 40(12): 3471-3477.
[12]	CUI Jianshuang, LIU Xiaochan, YANG Meihua, LI Wenyan. Meta-learning based optimization algorithm selection framework and its empirical study [J]. Journal of Computer Applications, 2017, 37(4): 1105-1110.
[13]	;;. Distributed association rules mining algorithm by sampling and meta-learning [J]. Journal of Computer Applications, 2006, 26(4): 872-874.