基于最小化背景判别性知识的小样本目标检测算法

张雅楠; 宋飞; 靳毅凡; 王晓明; 刘立祥; 李江梦

doi:10.16383/j.aas.c240341

基于最小化背景判别性知识的小样本目标检测算法

doi: 10.16383/j.aas.c240341 cstr: 32138.14.j.aas.c240341

张雅楠^{1, 2,},
宋飞^{1, 2,},
靳毅凡^{1, 2,},
王晓明^3,,
刘立祥^{1, 2,},
李江梦^1,

1.
中国科学院软件研究所北京 100190
2.
中国科学院大学北京 101408
3.
北京跟踪与通信技术研究所北京 100094

基金项目: 基础科研计划 (JCKY2022130C020)资助

详细信息

作者简介:
张雅楠：2025年获得中国科学院软件研究所博士学位. 主要研究方向为小样本学习, 目标检测. E-mail: yanan2018@iscas.ac.cn

宋飞：中国科学院软件研究所博士研究生. 2021年获得河南大学软件学院学士学位. 主要研究方向为多模态提示学习. E-mail: songfei2022@iscas.ac.cn

靳毅凡：中国科学院软件研究所在读博士研究生. 2020年获得中南大学计算机学院学士学位. 主要研究方向为深度学习, 图表示学习. E-mail: yifan2020@iscas.ac.cn

王晓明：北京跟踪与通信技术研究所. 2014年获得国防科技大学博士学位. 研究方向为天基智能化. E-mail: wxm_01@bittt.cn

刘立祥：中国科学院软件研究所研究员. 2002年毕业于上海交通大学电子工程学院, 获博士学位. 主要从事大型综合信息系统智能组网协议、大规模一体化系统地面验证以及复杂信息系统体系结构方面的研究. E-mail: lixiang@iscas.ac.cn

李江梦：中国科学院软件研究所助理研究员. 2023年在中国科学院软件研究所获得博士学位. 主要研究方向为可信多模态信息融合. 本文通信作者. E-mail: jiangmeng2019@iscas.ac.cn

计量
- 文章访问数: 37
- HTML全文浏览量: 40
- 被引次数: 0
出版历程
- 收稿日期: 2024-06-12
- 录用日期: 2025-02-14
- 网络出版日期: 2025-06-19

Minimizing Background Discriminative Knowledge for Few-shot Object Detection

ZHANG Ya-Nan^{1, 2
,},
SONG Fei^{1, 2
,},
JIN Yi-Fan^{1, 2
,},
WANG Xiao-Ming^3
,,
LIU Li-Xiang^{1, 2
,},
LI Jiang-Meng^1
,

1.
Institute of Software Chinese Academy of Sciences, Beijing 100190
2.
University of Chinese Academy of Sciences, Beijing 101408
3.
Beijing Institute of Tracking and Telecommunications Technology, Beijing 100094

Funds: Supported by the Fundamental Research Program, China (JCKY2022130C020)

More Information

Author Bio:
ZHANG Ya-Nan　Received her Ph.D. degree from the Institute of Software, Chinese Academy of Sciences in 2025. Her research interest covers few-shot learning and object detection

SONG Fei　Ph.D. candidate at the Institute of Software, Chinese Academy of Sciences. Received her bachelor degree from the Software College of Henan University in 2021. Her main research focus is on multimodal prompt learning

JIN Yi-Fan　Ph.D. candidate at the Institute of Software, Chinese Academy of Sciences. Received her bachelor degree from Central South University in 2020. Her main research focuses on deep learning and graph representation learning

WANG Xiao-Ming　Beijing Institute of Tracking and Telecommunications Technology. Received his Ph.D. degree from National University of Defense Technology in 2014. His research focuses on space-based intelligent systems

LIU Li-Xiang　Professor of the Institute of software of the Chinese Academy of Sciences (ISCAS). Graduated from the Shanghai Jiaotong University in 2002 with a doctor's degree. His research interests include intelligent networking of space-based integrated information systems, ground verification of large-scale network integrated system and complex information system architecture

LI Jiang-Meng　Assistant researcher at the Institute of Software, Chinese Academy of Sciences. Received his Ph.D. degree from the Institute of Software, Chinese Academy of Sciences in 2023. His primary research focuses on trustworthy multimodal information fusion. Corresponding author of this paper

摘要

摘要: 在小样本目标检测领域, “训练和微调”两阶段表征学习范式因学习策略简单, 应用广泛. 然而, 通过探索性实验发现, 基于该范式的模型容易将新类别实例错误地分类为背景类, 从而降低了对新类的识别能力. 为了解决这一问题, 提出构造一个正则化分类器, 并使用“最小化背景判别性知识的调节器 (Background discriminative knowledge minimizing regulator, BDKMR)”来引导分类器训练. BDKMR通过“最小化背景判别性知识的交叉$l_p $正则项(Background discriminative knowledge minimizing cross-$l_p $ regularization, BMCR)”显式地减少背景判别性知识对构建新类分类器的干扰, 并利用“权重范数管理器(Weight norm manager, WNM)”调节分类器中各类别的权重范数, 以提高模型对新类别的关注度, 同时降低其对背景类别的偏好. 此外, 考虑到 BDKMR 可能改变特征空间分布, 提出“分类器解耦模块(Decoupled box classifier, DBC)”, 以调控模型微调过程中正则化分类器对特征提取器学习的影响. 多个数据集上的实验结果表明, 所提出的方法能够有效减少模型对新类实例的错误分类, 进而显著提升对新类的检测性能.
- 小样本学习 /
- 小样本目标检测 /
- 正则项 /
- 判别性知识
Abstract: In the field of few-shot object detection, the “training and fine-tuning” two-stage representation learning framework is widely used due to the simplicity of its learning strategy. However, through exploratory experiments, we demonstrate that this learning paradigm is prone to misclassify novel instances as background instances, which hinders the ability of the model to recognize novel object instances. To address this issue, we propose the background discriminative knowledge minimizing regulator (BDKMR) to guide the classifier training. BDKMR explicitly reduces the effect of background discriminative knowledge on the classifier for novel categories by employing the background discriminative knowledge minimizing cross-$l_p $ regularization (BMCR). Moreover, BDKMR enhances the attention of the model on novel categories while alleviating its bias toward the background category through weight norm manager (WNM). Additionally, considering that BDKMR can alter the feature space distribution, the decoupled box classifier (DBC) module is introduced to adjust the impact of the regulator on the feature extractor during the fine-tuning stage. Experimental results on multiple datasets validate that the proposed method effectively reduces the misclassification of novel object instances and improves the performance of novel categories.
- Few-shot learning /
- few-shot object detection /
- regularization /
- discriminative knowledge

HTML全文

图 1 分类错误的统计结果

Fig. 1 Exploratory experiments of classification errors

下载: 全尺寸图片幻灯片

图 2 权重和特征沿维度的分布

Fig. 2 Distributions of weights and features along dimensions

下载: 全尺寸图片幻灯片

图 3 新类的权重范数平均值和正权重范数平均值随训练迭代次数的变化

Fig. 3 The variation of the average norm of weights and the average norm of positive weights for novel classes with respect to the number of training iterations

下载: 全尺寸图片幻灯片

图 4 不同类别权重范数与正权重范数的比较

Fig. 4 Comparisons of weight norms and positive weight norms between different categories

下载: 全尺寸图片幻灯片

图 5 前景类原型与背景类原型的相似度分析

Fig. 5 Analysis of the similarity between prototypes of foreground classes and the background

下载: 全尺寸图片幻灯片

图 6 前景类别与背景类别的分类器权重相似度分析

Fig. 6 Similarity analysis of classifier weights for foreground and background categories

下载: 全尺寸图片幻灯片

图 7 BDKMR方法概述图

Fig. 7 Illustration of BDKMR

下载: 全尺寸图片幻灯片

图 8 不同样本数设置下新类与背景类权重的相似度比较

Fig. 8 Comparison of weight similarity between novel and background categories under different shot settings

下载: 全尺寸图片幻灯片

图 9 DeFRCN和BDKMR的分类结果对比

Fig. 9 Comparison of classification results between DeFRCN and BDKMR

下载: 全尺寸图片幻灯片

图 10 BDKMR相较于DeFRCN的新类-背景类权重相似度变化

Fig. 10 Changes of weight similarity between novel and background classes: BDKMR compared to DeFRCN

下载: 全尺寸图片幻灯片

图 11 BDKMR和DeFRCN检测结果的可视化对比

Fig. 11 Visual comparison of detection results between BDKMR and DeFRCN

下载: 全尺寸图片幻灯片

表 1 COCO数据集中新类的小样本检测性能(%)

Table 1 FSOD results of novel categories on COCO (%)

方法	样本数
方法	1	2	3	5	10	30
FSRW^[22]	—	—	—	—	5.6	9.1
TFA^[14]	3.4	4.6	6.6	8.3	10.0	13.7
MPSR^[65]	2.3	3.5	5.2	6.7	9.8	14.1
FSCE^[25]	—	—	—	—	11.9	16.4
SRR-FSD^[16]	—	—	—	—	11.3	14.7
DeFRCN^[39]	6.5	11.8	13.4	15.3	18.6	22.5
FCT^[35]	5.6	7.9	11.1	14.0	17.1	21.4
BDKMR (本文)	9.8	13.3	13.8	15.5	18.1	21.6
注: 加粗字体表示各列最优结果

下载: 导出CSV

表 2 VOC数据集中新类的小样本检测性能(%)

Table 2 FSOD results of the novel categories on VOC (%)

方法	新类分割1					新类分割2					新类分割3
方法	1	2	3	5	10	1	2	3	5	10	1	2	3	5	10
TFA^[14]	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^[65]	41.7	42.5	51.4	55.2	61.8	24.4	29.3	39.2	39.9	47.8	35.6	41.8	42.3	48.0	49.7
FSCE^[25]	44.2	43.8	51.4	61.9	63.4	27.3	29.5	43.5	44.2	50.2	37.2	41.9	47.5	54.6	58.5
SRR-FSD^[16]	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*^[39]	55.1	61.9	64.9	65.8	66.2	33.8	45.1	46.1	53.2	52.3	51.0	56.6	55.6	59.7	61.9
Meta Faster R-CNN^[41]	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
FCT^[35]	49.9	57.1	57.9	63.2	67.1	27.6	34.5	43.7	49.2	51.2	39.5	54.7	52.3	57.0	58.7
Pseudo-Labelling^[40]	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
ICPE^[37]	54.3	59.5	62.4	65.7	66.2	33.5	40.1	48.7	51.7	52.5	50.9	53.1	55.3	60.6	60.1
$\sigma$-ADP^[36]	52.3	55.5	63.1	65.9	66.7	42.7	45.8	48.7	54.8	56.3	47.8	51.8	56.8	60.3	62.4
BDKMR (本文)	58.3	65.1	67.2	67.8	66.6	37.6	47.2	48.9	53.8	52.3	55.2	58.8	57.8	60.9	62.5

下载: 导出CSV

表 4 在VOC数据集的30个随机训练集上的实验结果(%)

Table 4 FSOD results on VOC over 30 random samples (%)

方法	新类分割1					新类分割2					新类分割3
方法	1	2	3	5	10	1	2	3	5	10	1	2	3	5	10
FRCN+ft-full^[33]	9.9	15.6	21.6	28.0	35.6	9.4	13.8	17.4	21.9	29.8	8.1	13.9	19.0	23.9	31.0
Xiao et al^[15]	24.2	35.3	42.2	49.1	57.4	21.6	24.6	31.9	37.0	45.7	21.2	30.0	37.2	43.8	49.6
TFA^[14]	25.3	36.4	42.1	47.9	52.8	18.3	27.5	30.9	34.1	39.5	17.9	27.2	34.3	40.8	45.6
FSCE^[25]	32.9	44.0	46.8	52.9	59.7	23.7	30.6	38.4	43.0	48.5	22.6	33.4	39.5	47.3	54.0
DeFRCN*^[39]	39.3	50.9	55.3	61.8	65.3	27.4	36.8	40.4	45.1	50.8	35.0	45.1	50.2	55.7	58.9
DCNet^[34]	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
FCT^[35]	38.5	49.6	53.5	59.8	64.3	25.9	34.2	40.1	44.9	47.4	34.7	43.9	49.3	53.1	56.3
$\sigma$-ADP^[36]	35.9	40.3	49.8	56.8	65.1	25.6	30.3	41.7	41.8	50.3	33.9	35.6	43.5	47.1	55.9
BDKMR (本文)	43.6	54.3	58.1	63.1	66.4	30.1	38.2	41.5	45.7	51.1	39.9	48.3	52.4	57.5	59.8
BDKMR (本文)	$\pm$3.2	$\pm$2.0	$\pm$1.8	$\pm$1.1	$\pm$1.0	$\pm$2.7	$\pm$2.1	$\pm$1.7	$\pm$1.9	$\pm$1.0	$\pm$3.2	$\pm$2.2	$\pm$1.5	$\pm$1.1	$\pm$1.1

下载: 导出CSV

表 3 在COCO数据集10个随机训练集上的实验结果(%)

Table 3 Experimental results over 10 random samples on COCO (%)

样本数	方法	新类
样本数	方法	nAP	nAP$_{50}$	nAP$_{75}$
1	FRCN+ft-full^[33]	1.7$\pm$0.2	3.3$\pm$0.3	1.6$\pm$0.2
	TFA^[14]	1.9$\pm$0.4	3.8$\pm$0.6	1.7$\pm$0.5
	DeFRCN^[39]	4.8$\pm$0.6	9.5$\pm$0.9	4.4$\pm$0.8
	BDKMR (本文)	7.3$\pm$0.6	14.7$\pm$0.9	6.5$\pm$0.8
2	FRCN+ft-full^[33]	3.1$\pm$0.3	6.1$\pm$0.6	2.9$\pm$0.3
	TFA^[14]	3.9$\pm$0.4	7.8$\pm$0.7	3.6$\pm$0.6
	DeFRCN^[39]	8.5$\pm$0.9	16.3$\pm$1.4	7.8$\pm$1.1
	BDKMR (本文)	10.7$\pm$0.7	20.6$\pm$1.1	9.9$\pm$0.9
3	FRCN+ft-full^[33]	3.7$\pm$0.4	7.1$\pm$0.8	3.5$\pm$0.4
	TFA^[14]	5.1$\pm$0.6	9.9$\pm$0.9	4.8$\pm$0.6
	DeFRCN^[39]	10.7$\pm$0.7	20.0$\pm$1.2	10.3$\pm$0.8
	BDKMR (本文)	12.4$\pm$0.4	23.2$\pm$0.9	11.8$\pm$0.5
5	FRCN+ft-full^[33]	4.6$\pm$0.5	8.7$\pm$1.0	4.4$\pm$0.6
	TFA^[14]	7.0$\pm$0.7	13.3$\pm$1.2	6.5$\pm$0.7
	DeFRCN^[39]	13.5$\pm$0.6	24.7$\pm$1.1	13.0$\pm$0.6
	BDKMR (本文)	14.3$\pm$0.5	26.5$\pm$1.2	13.8$\pm$0.5
10	FRCN+ft-full^[33]	5.5$\pm$0.9	10.0$\pm$1.6	5.5$\pm$0.9
	TFA^[14]	9.1$\pm$0.5	17.1$\pm$1.1	8.8$\pm$0.5
	DeFRCN^[39]	16.7$\pm$0.6	29.6$\pm$1.3	16.7$\pm$0.4
	BDKMR (本文)	16.9$\pm$0.5	30.1$\pm$1.1	16.8$\pm$0.6
30	FRCN+ft-full^[33]	7.4$\pm$1.1	13.1$\pm$2.1	7.4$\pm$1.0
	TFA^[14]	12.1$\pm$0.4	22.0$\pm$0.7	12.0$\pm$0.5
	DeFRCN^[39]	21.0$\pm$0.4	36.7$\pm$0.8	21.4$\pm$0.4
	BDKMR (本文)	19.7$\pm$0.5	34.4$\pm$1.0	20.2$\pm$0.4

下载: 导出CSV

表 5 基于COCO数据集的分类器消融实验结果(%)

Table 5 Ablative experimental results in the classifiers on COCO (%)

分类器			nAP						nAP$_{50}$
BMCR	PWE	WN	1	2	3	5	10	30	1	2	3	5	10	30
$\times$	$\times$	$\times$	7.3	10.9	12.5	14.1	17.1	20.0	13.4	20.9	23.8	27.7	32.9	37.0
$\times$	$\times$	$\checkmark$	8.0	11.4	12.7	14.1	16.9	19.7	14.4	21.5	24.1	27.4	32.6	36.6
$\checkmark$	$\times$	$\checkmark$	8.4	11.7	13.0	14.3	17.2	19.8	15.3	22.0	24.9	28.0	32.9	36.7
$\times$	$\checkmark$	$\checkmark$	8.3	11.6	12.6	13.9	16.8	19.4	15.2	22.0	24.1	27.4	32.5	36.3
$\checkmark$	$\checkmark$	$\checkmark$	8.9	12.1	12.9	14.3	17.0	19.5	17.0	23.1	25.1	28.2	32.9	36.2

下载: 导出CSV

表 6 基于COCO数据集的特征提取器消融实验结果(%)

Table 6 Comparison results of different fine-tuning methods on COCO (%)

BDKMR	FE		nAP						nAP$_{50}$
BDKMR	微调	DBC	1	2	3	5	10	30	1	2	3	5	10	30
$\times$	$\checkmark$	$\times$	6.5	11.8	13.4	15.3	18.6	22.5	11.0	20.6	24.3	28.4	34.6	39.9
$\checkmark$	$\checkmark$	$\times$	8.7	12.1	13.0	14.7	17.4	20.8	15.8	22.4	24.7	28.7	33.2	37.7
$\checkmark$	$\checkmark$	$\checkmark$	9.8	13.3	13.8	15.5	18.1	21.6	17.9	24.7	26.3	30.4	34.4	39.0

下载: 导出CSV

表 7 COCO数据集上不同$\alpha $值下的实验结果(%)

Table 7 Experimental results for different $\alpha $ values on COCO (%)

$\alpha$	样本数				合计$\Delta$
$\alpha$	1	2	3	10	合计$\Delta$
0	8.0	11.4	12.7	16.9	—
1	8.1	11.5	12.7	17.0	+0.3
10	8.3	11.5	12.6	16.9	+0.3
30	8.4	11.8	13.0	16.9	+1.1
50	8.4	11.7	13.0	17.2	+1.3
70	8.3	11.4	12.8	17.0	+0.5
90	8.3	11.6	13.0	17.0	+0.9
5 000	7.3	10.3	11.5	15.7	−4.2

下载: 导出CSV

表 8 COCO数据集上不同$\beta $值下的实验结果(%)

Table 8 Experimental results for different $\beta $values on COCO (%)

$\beta$	样本数				合计$\Delta$
$\beta$	1	2	3	10	合计$\Delta$
0.0	8.4	11.7	13.0	17.2	—
0.1	8.7	11.9	13.1	17.0	+0.4
0.2	8.9	12.1	12.9	17.0	+0.6
0.4	9.1	11.9	12.7	16.7	+0.1
0.8	9.0	11.7	12.4	16.3	−0.9

下载: 导出CSV

表 9 COCO数据集上不同p值下的实验结果(%)

Table 9 Experimental results for different $p $ values on COCO (%)

$p$	样本数				合计$\Delta$
$p$	1	2	3	10	合计$\Delta$
固定特征提取器参数
—	7.3	10.9	12.5	17.1	—
1	8.7	11.7	12.8	17.0	+2.4
2	8.9	12.1	12.9	17.0	+3.1
3	8.8	12.0	13.0	16.7	+2.7
4	8.9	12.1	12.8	16.6	+2.4
5	8.7	12.1	12.8	16.6	+2.4
更新特征提取器参数
—	6.5	11.8	13.4	18.6	—
1	9.6	13.0	13.8	18.0	+4.1
2	9.8	13.3	13.8	18.1	+4.7
3	10.0	13.2	14.1	18.0	+5.0
4	9.2	13.0	13.5	17.5	+2.9
5	9.1	12.7	13.8	17.6	+2.8

下载: 导出CSV

表 10 COCO数据集上两种PWE实现方式的比较(%)

Table 10 Comparisons of two implementations of PWE on the COCO dataset (%)

APWE	CPWE	$\gamma$	样本数				合计$\Delta$
APWE	CPWE	$\gamma$	1	2	3	10	合计$\Delta$
$\times$	$\times$	—	8.0	11.4	12.7	16.9	—
$\checkmark$	$\times$	—	8.3	11.6	12.6	16.8	+0.3
$\times$	$\checkmark$	1.00	8.3	11.2	12.2	16.2	−1.1
$\times$	$\checkmark$	0.95	8.4	11.6	12.6	16.7	+0.3
$\times$	$\checkmark$	0.90	8.2	11.6	12.7	16.8	+0.3
$\times$	$\checkmark$	0.85	8.2	11.5	12.7	17.0	+0.4
$\times$	$\checkmark$	0.80	8.1	11.6	12.8	16.8	+0.3

下载: 导出CSV

表 11 将BDKMR插入到其他基线方法的模型性能(%)

Table 11 Model performance by integrating BDKMR into other baseline methods (%)

方法	集成 BDKMR	样本数
方法	集成 BDKMR	1	2	3	5	10
TFA-fc^[14]	$\times$	41.5*	38.0*	43.7*	54.7*	56.0*
TFA-fc^[14]	$ \checkmark$	44.9	41.9	44.4	54.9	55.4
TFA-cosine^[14]	$\times$	46.3*	41.5*	45.0*	53.8*	54.4*
TFA-cosine^[14]	$ \checkmark$	47.0	42.7	45.4	54.7	54.9
DeFRCN^[39]	$\times$	55.1*	61.9*	64.9*	65.8*	66.2*
DeFRCN^[39]	$ \checkmark$	58.3	65.1	67.2	67.8	66.6

下载: 导出CSV

表 12 基于VOC数据集新类分割1的全部实验结果(%)

Table 12 Full results on the novel set 1 of VOC (%)

方法	1-shot			2-shot			3-shot			5-shot			10-shot
方法	mAP	bAP	nAP	mAP	bAP	nAP	mAP	bAP	nAP	mAP	bAP	nAP	mAP	bAP	nAP
MPSR^[65]	56.8	—	41.7	60.4	—	42.5	62.8	—	51.4	66.1	—	55.2	69.0	—	61.8
TFA w/ fc^[14]	69.3	80.2	36.8	66.9	79.5	29.1	70.3	79.2	43.6	73.4	79.2	55.7	73.2	78.6	57.0
TFA w/ cosine^[14]	69.7	79.6	39.8	68.2	78.9	36.1	70.5	79.1	44.7	73.4	79.3	55.7	72.8	78.4	56.0
Retentive R-CNN^[45]	71.3	—	42.4	72.3	—	45.8	72.1	—	45.9	74.0	—	53.7	74.6	—	56.1
DeFRCN*^[39]	72.4	78.1	55.1	73.1	76.9	61.9	73.5	76.4	64.9	74.3	77.2	65.8	74.1	76.7	66.2
BDKMR	73.3	78.3	58.3	74.0	77.0	65.1	74.7	77.1	67.2	75.0	77.4	67.8	74.5	77.1	66.6

下载: 导出CSV

表 13 模型复杂度对比分析

Table 13 Comparative analysis of model complexity

模型	训练		测试
模型	时间	参数量	时间	参数量
DeFRCN^[39]	0.7s/iter	约5 237.5万	0.7s/img	约5 237.5万
BDKMR	0.7s/iter	约5 241.8万	0.7s/img	约5 237.5万

下载: 导出CSV

参考文献(66)

[1]	Redmon J, Divvala S, Girshick R, Farhadi A. You only look once: Unified, real-time object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Las Vegas, Nevada, USA: 2016. 779−788
[2]	Lin T Y, Goyal P, Girshick R, He K, Dollár P. Focal loss for dense object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. Venice, Italy: IEEE 2017. 2999−3007
[3]	Ren S, He K, Girshick R, Sun J. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017, 39(6): 1137−1149 doi: 10.1109/TPAMI.2016.2577031
[4]	He K, Gkioxari G, Dollár P, Girshick R. Mask r-cnn. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. Venice, Italy: IEEE 2017. 2980−2988
[5]	Tian Z, Shen C, Chen H, He T. FCOS: a simple and strong anchor-free object detector. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022, 44(4): 1922−1933
[6]	Zhou X, Wang D, Krähenbühl P. Objects as points. arXiv preprint arXiv: 1904.07850, 2019.
[7]	刘小波, 肖肖, 王凌, 蔡之华, 龚鑫, 郑可心. 基于无锚框的目标检测方法及其在复杂场景下的应用进展. 自动化学报, 2023, 49(7): 1369−1392 Liu Xiao-Bo, Xiao Xiao, Wang Ling, Cai Zhi-Hua, Gong Xin, Zheng Ke-Xin. Anchor-free based object detection methods and its application progress in complex scenes. Acta Automatica Sinica, 2023, 49(7): 1369−1392
[8]	Chen S, Sun P, Song Y, Luo P. Diffusiondet: Diffusion model for object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. Paris, France: IEEE 2023. 19773−19786
[9]	Zhao Y, Lv W, Xu S, Wei J, Wang G, Dang Q, et al. Detrs beat yolos on real-time object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Seattle, WA, USA: 2024. 16965−16974
[10]	Wang A, Chen H, Liu L, Chen K, Lin Z, Han J, et al. Yolov10: Real-time end-to-end object detection. In: Proceedings of Advances in Neural Information Processing Systems. Vancouver, BC, Canada: 2024. 107984−108011
[11]	Pachetti E, Colantonio S. A systematic review of few-shot learning in medical imaging. Artificial Intelligence in Medicine, DOI: 10.1016/J.ARTMED.2024.102949
[12]	Zhou Z, Zhao L, Ji K, Kuang G. A domain adaptive few-shot SAR ship detection algorithm driven by the latent similarity between optical and SAR images. IEEE Transactions on Geoscience and Remote Sensing, DOI: 10.1109/TGRS.2024.3421512
[13]	Chen H, Wang Y, Wang G, Qiao Y. Lstd: A low-shot transfer detector for object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. New Orleans, Louisiana, USA: 2018. 2836−2843
[14]	Wang X, Huang T E, Gonzalezand J, Darrell T, Yu F. Frustratingly simple few-shot object detection. In: Proceedings of the 37nd International Conference on Machine Learning. Virtual Conference: 2020. 9919−9928
[15]	Xiao Y, Lepetit V, Marlet R. Few-shot object detection and viewpoint estimation for objects in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022, 45(3): 3090−3106
[16]	Zhu C, Chen F, Ahmed U, Shen Z, Savvides M. Semantic relation reasoning for shot-stable few-shot object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Virtual Conference: 2021. 8782−8791
[17]	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
[18]	王多瑞, 杜杨, 董兰芳, 胡卫明, 李兵. 基于特征变换和度量网络的小样本学习算法. 自动化学报, 2023, 50(7): 1305−1314 Wang Duo-Rui, Du Yang, Dong Lan-Fang, Hu Wei-Ming, Li Bing. Metric based feature transformation networks for few-shot learning. Acta Automatica Sinica, 2023, 50(7): 1305−1314
[19]	Zhang R, Tan J, Cao Z, Xu L, Liu Y, Si L, et al. Part-aware correlation networks for few-shot learning. IEEE Transactions on Multimedia, DOI: 10.1109/TMM.2024.3394681
[20]	Zhang H, Xu J, Jiang S, He Z. Simple semantic-aided few-shot learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Seattle, WA, USA: 2024. 28588−28597
[21]	Xin Z, Chen S, Wu T, Shao Y, Ding W, You X. Few-shot object detection: Research advances and challenges. Information Fusion, DOI: 10.1016/J.INFFUS.2024.102307
[22]	Kang B, Liu Z, Wang X, Yu F, Feng J, Darrell T. Few-shot object detection via feature reweighting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. Seoul, South Korea: IEEE 2019. 8419−8428
[23]	Wang Y X, Ramanan D, Hebert M. Meta-learning to detect rare objects. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. Seoul, South Korea: IEEE 2019. 9924−9933
[24]	Wang Z, Yang B, Yue H, Ma Z. Fine-grained prototypes distillation for few-shot object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. Vancouver, Canada: 2024. 5859−5866
[25]	Sun B, Li B, Cai S, Yuan Y, Zhang C. Fsce: Few-shot object detection via contrastive proposal encoding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Virtual Conference: 2021. 7352−7362
[26]	Ma J, Niu Y, Xu J, Huang S, Han G, Chang S F. Digeo: Discriminative geometry-aware learning for generalized few-shot object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Vancouver, BC, Canada: 2023. 3208−3218
[27]	Xin Z, Wu T, Chen S, Zou Y, Shao L, You X. ECEA: Extensible co-existing attention for few-shot object detection. IEEE Transactions on Image Processing, DOI: 10.1109/TIP.2024.3411771
[28]	Lin T Y, Maire M, Belongie S J, Hays J, Perona P, Ramanan D, et al. Microsoft coco: common objects in context. In: Proceedings of European Conference on Computer Vision. Zurich, Switzerland: 2014. 740−755
[29]	Finn C, Abbeel P, Levine S. Model-agnostic meta-learning for fast adaptation of deep networks. In: Proceedings of the 34nd International Conference on Machine Learning. Sydney, Australia: 2017. 1126−1135
[30]	Li Z, Zhou F, Chen F, Li H. Meta-sgd: Learning to learn quickly for few-shot learning. arXiv preprint arXiv: 1707.09835, 2017.
[31]	Nichol A, Achiam J, Schulman J. On first-order meta-learning algorithms. arXiv preprint arXiv: 1803.02999, 2018.
[32]	Luo X, Wu H, Zhang J, Gao L, Xu J, Song J. A closer look at few-shot classification again. In: Proceedings of the International Conference on Machine Learning. Honolulu, Hawaii, USA: 2023. 23103−23123
[33]	Yan X, Chen Z, Xu A, Wang X, Liang X, Lin L. Meta r-cnn: Towards general solver for instance-level low-shot learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. Seoul, South Korea: IEEE 2019. 9576−9585
[34]	Hu H, Bai S, Li A, Cui J, Wang L. Dense relation distillation with context-aware aggregation for few-shot object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Virtual Conference: 2021. 10185−10194
[35]	Han G, Ma J, Huang S, Chen L, Chang S F. Few-shot object detection with fully cross-transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. New Orleans, Louisiana, USA: 2022. 5321−5330
[36]	Du J, Zhang S, Chen Q, Le H, Sun Y, Ni Y, et al. σ-adaptive decoupled prototype for few-shot object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. Paris, France: IEEE 2023. 18904−18914
[37]	Lu X, Diao W, Mao Y, Li J, Wang P, Sun X, et al. Breaking immutable: Information-coupled prototype elaboration for few-shot object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. Washington, DC, USA: 2023. 1844−1852
[38]	Zhao X, Zou X, Wu Y. Morphable detector for object detection on demand. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. Virtual Conference: IEEE 2021. 4751−4760
[39]	Qiao L, Zhao Y, Li Z, Qiu X, Wu J, Zhang C. Defrcn: Decoupled faster r-cnn for few-shot object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. Virtual Conference: IEEE 2021. 8661−8670
[40]	Kaul P, Xie W, Zisserman A. Label, verify, correct: A simple few shot object detection method. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. New Orleans, Louisiana, USA: 2022. 14237−14247
[41]	Han G, Huang S, Ma J, He Y, Chang S F. Meta faster r-cnn: Towards accurate few-shot object detection with attentive feature alignment. In: Proceedings of the AAAI Conference on Artificial Intelligence. Vancouver, BC, Canada: 2022. 780−789
[42]	Zhang G, Luo Z, Cui K, Lu S, Xing E P. Meta-detr: Image-level few-shot detection with inter-class correlation exploitation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022, 45(11): 12832−12843
[43]	Vaswani A, Shazeer N, Parmar N, Uszkoreit J, Jones L, Gomez A N, et al. Attention is all you need. In: Proceedings of Advances in Neural Information Processing Systems. Long Beach, CA, USA: 2017. 5998−6008
[44]	Carion N, Massa F, Synnaeve G, Usunier N, Kirillov A, Zagoruyko S. End-to-end object detection with transformers. In: Proceedings of European Conference on Computer Vision. Glasgow, UK: 2020. 213−229
[45]	Fan Z, Ma Y, Li Z, Sun J. Generalized few-shot object detection without forgetting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Virtual Conference: 2021. 4527−4536
[46]	Li J, Zhang Y, Qiang W, Si L, Jiao C, Hu X, et al. Disentangle and remerge: interventional knowledge distillation for few-shot object detection from a conditional causal perspective. In: Proceedings of the AAAI Conference on Artificial Intelligence. Washington, DC, USA: 2023. 1323−1333
[47]	Li A, Li Z. Transformation invariant few-shot object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Virtual Conference: 2021. 3094−3102
[48]	Xu J, Le H, Samaras D. Generating features with increased crop-related diversity for few-shot object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Vancouver, BC, Canada: 2023. 19713−19722
[49]	Salimans T, Kingma D P. Weight normalization: A simple reparameterization to accelerate training of deep neural networks. In: Proceedings of Advances in Neural Information Processing Systems. Barcelona, Spain: 2016. 29
[50]	Liu W, Wen Y, Yu Z, Li M, Raj B, Song L. Sphereface: Deep hypersphere embedding for face recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Honolulu, HI, USA: IEEE 2017. 6738−6746
[51]	Liu Y, Li H, Wang X. Learning deep features via congenerous cosine loss for person recognition. arXiv preprint arXiv: 1702.06890, 2017.
[52]	Hasnat M A, Bohné J, Milgram J, Gentric S, Chen L. von mises-fisher mixture model-based deep learning: Application to face verification. arXiv preprint arXiv: 1706.04264, 2017.
[53]	Wang F, Xiang X, Cheng J, Yuille A L. Normface: L₂ hypersphere embedding for face verification. In: Proceedings of the 25th ACM International Conference on Multimedia. Mountain View, CA, USA: 2017. 1041−1049
[54]	Chen W Y, Liu Y C, Kira Z, Wang Y C F, Huang J B. A closer look at few-shot classification. In: Proceedings of the 6th International Conference on Learning Representations. New Orleans, Louisiana, USA: ICLR, 2019.
[55]	Tibshirani R. Regression shrinkage and selection via the lasso. Journal of the Royal Statistical Society Series B: Statistical Methodology, 1996, 58(1): 267−288 doi: 10.1111/j.2517-6161.1996.tb02080.x
[56]	De M C, De V E, Rosasco L. Elastic-net regularization in learning theory. Journal of Complexity, 2009, 25(2): 201−230 doi: 10.1016/j.jco.2009.01.002
[57]	Liu S, Wei L, Lv S, Li M. Stability and Generalization of lp-Regularized Stochastic Learning for GCN. In: Proceedings of the International Joint Conferences on Artificial Intelligence. Macao, China: 2023. 5685−5693
[58]	Ribeiro A, Zachariah D, Bach F, Schön T. Regularization properties of adversarially-trained linear regression. In: Proceedings of Advances in Neural Information Processing Systems. New Orleans, Louisiana, USA: 2023. 36
[59]	Guo Y, Zhang L. One-shot face recognition by promoting underrepresented classes. arXiv preprint arXiv: 1707.05574, 2017.
[60]	Dang W, Yang Z, Dong W, Li X, Shi G. Inverse weight-balancing for deep long-tailed learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. Vancouver, Canada: 2024. 11713−11721
[61]	Snell J, Swersky K, Zemel R. Prototypical networks for few-shot learning. In: Proceedings of Advances in Neural Information Processing Systems. Long Beach, CA, USA: 2017. 4077−4087
[62]	Everingham M, Van Gool L, Williams C K I, Winn J M, Zisserman A. The pascal visual object classes (voc) challenge. International Journal of Computer Vision, 2010, 88(2): 303−338 doi: 10.1007/s11263-009-0275-4
[63]	He K, Zhang X, Ren S, Sun J. Deep residual learning for image recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Las Vegas, Nevada, USA: 2016. 770−778
[64]	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(3): 211−252 doi: 10.1007/s11263-015-0816-y
[65]	Wu J, Liu S, Huang D, Wang Y. Multi-scale positive sample refinement for few-shot object detection. In: Proceedings of European Conference on Computer Vision. Virtual Conference: 2020. 456−472
[66]	McCulloch J A, St. Pierre S R, Linka K, Kuhl E. On sparse regression, lp-regularization, and automated model discovery. International Journal for Numerical Methods in Engineering, 2024, 125(14): e7481 doi: 10.1002/nme.7481