多尺度选择金字塔网络的小样本目标检测算法

doi:10.3778/j.issn.1673-9418.2109081

摘要/Abstract

摘要：

目标检测是把图像中指定的目标检测出来,这一技术已经广泛运用于自动驾驶、人脸识别等领域,已成为国内外计算机视觉领域的一大研究热点。传统的目标检测往往需要大量标注的数据集,如何在只有少量带注释样本的情况下进行目标检测是一个挑战。针对此问题,提出了一种多尺度选择金字塔网络的小样本目标检测算法,使检测不再依赖于大规模标签数据集。首先,设计了一个用于小样本目标检测的多尺度选择金字塔网络,它由三个组件组成：上下文层注意力模块、特征尺度增强模块、特征尺度选择模块。然后,在RPN网络产生的RoI特征后采用最大池化和平均池化来提升特征之间的相关性,之后进行特征融合,并且采用特征减法来突出特征中的类别信息,在保持模型对样本参数稳定性的前提下提高了对新类参数的敏感度;最后,采用正交映射损失函数使模型在分类层前就约束特征,即使在少量样本情况下也能够很好地衡量特征间的相似性。

关键词: 目标检测, 小样本, 正交映射损失, 特征融合

Abstract:

Target detection is to detect the specified target in the image. This technology has been widely used in automatic driving, face recognition and other fields, and has become a major research hotspot in the field of computer vision at home and abroad. Traditional target detection often requires a large number of annotated datasets, so it is a challenge to detect targets with only a small number of annotated samples. To address this problem, this paper proposes a multi-scale selection pyramid network algorithm for small sample target detection so that detection no longer relies on large-scale labeled datasets. Firstly, this paper designs a multi-scale selection pyramid network for small sample target detection, which consists of three components: context layer attention module, feature scale enhancement module, and feature scale selection module. Secondly, this paper performs feature fusion after the RoI features generated by the RPN network using maximum pooling and average pooling to improve the correlation between features. This paper uses feature subtraction to highlight the category information in the features, which can improve the sensitivity to new class parameters while maintaining the stability of the model to the sample parameters. Finally, the orthogonal mapping loss function is used to constrain the features before the classification layer, which can well measure the similarity between features even in the case of a small number of samples.

Key words: target detection, small sample, orthogonal mapping loss, feature fusion

中图分类号:

TP391.4

彭豪, 李晓明. 多尺度选择金字塔网络的小样本目标检测算法[J]. 计算机科学与探索, 2022, 16(7): 1649-1660.

PENG Hao, LI Xiaoming. Multi-scale Selection Pyramid Networks for Small-Sample Target Detection Algorithms[J]. Journal of Frontiers of Computer Science and Technology, 2022, 16(7): 1649-1660.

图/表 13

图1 Meta R-CNN结构图

Fig.1 Meta R-CNN structure diagram

图2 提出的小样本目标检测算法的体系结构

Fig.2 Architecture of proposed small sample target detection algorithm

图3 不同层注意力热力图和不同模块互补作用分析

Fig.3 Heat map of attention on different layers and analysis of complementary effects of different modules

图4 多元特征融合示意图

Fig.4 Schematic diagram of multiple feature fusion

图5 特征减法示意图

Fig.5 Schematic diagram of feature subtraction

表1 VOC07测试集上的low-shot检测mAP

Table 1 Low-shot detection mAP on VOC07 test suite %

Methods	Novel-class split 1					Novel-class split 2					Novel-class split 3
Methods	$K = 1$	$K = 2$	$K = 3$	$K = 5$	$K = 10$	$K = 1$	$K = 2$	$K = 3$	$K = 5$	$K = 10$	$K = 1$	$K = 2$	$K = 3$	$K = 5$	$K = 10$
FSRW^[11]	14.5	15.3	26.9	32.4	46.8	15.7	16.8	21.4	31.8	39.3	21.1	25.9	28.4	40.6	45.7
MetaDet^[22]	18.8	19.4	28.6	35.2	49.0	20.6	23.5	27.5	31.7	42.5	20.5	23.4	29.7	42.1	44.5
FSOD-KT^[18]	27.6	40.8	46.5	55.3	56.2	19.2	26.5	37.0	38.8	41.4	29.5	30.8	38.5	42.7	45.6
TFA w/fc^[23]	36.8	35.85	41.6	55.8	56.9	18.3	27.1	31.5	35.4	38.2	27.8	33.4	40.5	48.4	50.9
TFA w/cos^[23]	39.7	36.9	44.3	54.2	56.5	23.4	26.4	32.8	34.9	38.0	30.4	35.1	41.7	49.5	48.4
FSDetView^[24]	24.5	35.5	42.9	48.5	54.3	21.5	27.6	31.8	37.5	44.9	21.4	30.2	35.8	37.7	49.5
MPSR^[15]	41.6	—	51.0	55.6	61.7	24.3	—	39.4	39.7	47.2	35.4	—	42.1	48.1	49.5
AFD-Net^[25]	33.2	39.5	51.6	55.2	60.6	24.8	29.7	40.5	44.2	48.1	27.3	35.1	43.5	47.6	51.2
CME^[26]	41.4	45.8	50.2	54.7	61.2	27.6	30.5	41.9	42.1	47.8	34.9	39.4	45.7	48.3	50.4
DCNet^[27]	33.8	37.2	42.8	51.3	59.4	23.4	24.4	30.6	36.8	46.4	32.0	34.6	39.8	42.7	49.6
Meta R-CNN^[3]	19.7	25.6	35.1	45.6	50.8	10.7	19.3	29.4	30.2	44.8	14.5	18.1	27.6	40.9	46.3
Ours	36.2	47.2	52.4	55.6	62.8	28.4	34.2	36.5	41.1	52.7	37.1	42.5	45.2	49.3	55.8

表1 VOC07测试集上的low-shot检测mAP

Table 1 Low-shot detection mAP on VOC07 test suite %

Methods	Novel-class split 1					Novel-class split 2					Novel-class split 3
Methods	$K = 1$	$K = 2$	$K = 3$	$K = 5$	$K = 10$	$K = 1$	$K = 2$	$K = 3$	$K = 5$	$K = 10$	$K = 1$	$K = 2$	$K = 3$	$K = 5$	$K = 10$
FSRW^[11]	14.5	15.3	26.9	32.4	46.8	15.7	16.8	21.4	31.8	39.3	21.1	25.9	28.4	40.6	45.7
MetaDet^[22]	18.8	19.4	28.6	35.2	49.0	20.6	23.5	27.5	31.7	42.5	20.5	23.4	29.7	42.1	44.5
FSOD-KT^[18]	27.6	40.8	46.5	55.3	56.2	19.2	26.5	37.0	38.8	41.4	29.5	30.8	38.5	42.7	45.6
TFA w/fc^[23]	36.8	35.85	41.6	55.8	56.9	18.3	27.1	31.5	35.4	38.2	27.8	33.4	40.5	48.4	50.9
TFA w/cos^[23]	39.7	36.9	44.3	54.2	56.5	23.4	26.4	32.8	34.9	38.0	30.4	35.1	41.7	49.5	48.4
FSDetView^[24]	24.5	35.5	42.9	48.5	54.3	21.5	27.6	31.8	37.5	44.9	21.4	30.2	35.8	37.7	49.5
MPSR^[15]	41.6	—	51.0	55.6	61.7	24.3	—	39.4	39.7	47.2	35.4	—	42.1	48.1	49.5
AFD-Net^[25]	33.2	39.5	51.6	55.2	60.6	24.8	29.7	40.5	44.2	48.1	27.3	35.1	43.5	47.6	51.2
CME^[26]	41.4	45.8	50.2	54.7	61.2	27.6	30.5	41.9	42.1	47.8	34.9	39.4	45.7	48.3	50.4
DCNet^[27]	33.8	37.2	42.8	51.3	59.4	23.4	24.4	30.6	36.8	46.4	32.0	34.6	39.8	42.7	49.6
Meta R-CNN^[3]	19.7	25.6	35.1	45.6	50.8	10.7	19.3	29.4	30.2	44.8	14.5	18.1	27.6	40.9	46.3
Ours	36.2	47.2	52.4	55.6	62.8	28.4	34.2	36.5	41.1	52.7	37.1	42.5	45.2	49.3	55.8

表2 VOC07测试集上的 AP 和 mAP

Table 2 AP and mAP on VOC07 test suite %

Shot	Methods	Novel-class split 1						Novel-class split 2						Novel-class split 3
Shot	Methods	bird	bus	cow	mobike	sofa	mean	aero	bottle	cow	horse	sofa	mean	boat	cat	mobike	sheep	sofa	mean
$K = 1$	FSRW^[11]	13.2	15.5	30.7	35.2	7.9	20.5	11.7	9.1	15.2	23.9	18.4	15.7	10.5	44.0	18.4	18.3	5.2	19.3
	MetaDet^[22]	6.2	32.5	15.4	35.6	0.7	18.1	23.5	0.9	23.8	4.8	0.9	10.8	4.6	31.4	11.5	11.8	0.2	11.9
	MPSR^[15]	33.8	41.9	57.2	53.0	21.4	41.5	21.4	9.5	35.9	30.6	25.3	24.5	14.3	47.8	35.4	32.4	22.4	30.5
	Ours	48.6	39.4	44.6	53.5	21.5	41.5	31.8	7.6	27.6	39.4	38.7	29.0	19.7	55.9	17.2	18.2	43.9	31.0
$K = 2$	FSRW^[11]	21.3	12.4	16.7	17.2	9.3	15.4	28.5	0.9	27.3	0.0	19.2	15.2	5.2	46.5	18.9	26.7	12.0	21.9
	MetaDet^[22]	35.2	47.5	54.8	32.2	0.4	34.0	12.3	1.0	44.1	52.9	0.8	22.2	16.5	25.8	19.5	0.9	18.5	16.2
	MPSR^[15]	38.5	16.4	56.3	57.3	32.4	40.2	36.4	9.1	46.8	27.5	34.2	30.8	18.7	45.3	59.7	49.2	32.8	41.1
	Ours	48.3	55.7	52.0	59.5	29.8	49.1	36.5	9.7	51.2	55.7	17.9	34.2	15.9	69.7	59.9	41.4	37.2	44.8
$K = 3$	FSRW^[11]	26.2	19.4	20.5	20.5	27.0	27.7	29.4	4.9	34.5	6.7	37.6	22.6	11.3	39.5	20.7	23.1	33.6	25.6
	MetaDet^[22]	32.4	46.1	53.85	38.4	10.7	36.3	26.8	0.3	50.6	53.8	18.9	30.1	13.6	39.3	32.8	38.4	10.5	26.9
	MPSR^[15]	35.1	60.7	61.4	61.3	43.6	52.5	47.6	9.0	47.3	45.2	40.5	37.9	14.9	62.8	57.9	37.5	42.1	43.0
	Ours	59.4	67.8	69.3	64.9	45.2	61.3	45.4	15.6	47.6	46.6	38.9	38.8	24.2	65.1	55.9	52.8	36.8	47.0
$K = 5$	FSRW^[11]	31.6	22.5	39.6	40.0	37.1	34.2	33.1	9.6	38.4	25.9	44.0	30.2	14.5	57.6	50.2	38.6	41.5	40.5
	MetaDet^[22]	35.2	47.5	54.8	55.1	34.8	45.5	28.3	1.7	50.6	54.9	38.4	34.8	16.5	45.8	53.1	40.3	48.5	40.8
	MPSR^[15]	37.9	65.5	55.1	68.6	47.5	54.9	47.6	10.7	45.1	45.8	47.6	39.4	20.7	56.8	68.1	48.7	45.6	48.0
	Ours	58.3	65.7	43.0	66.7	41.5	55.0	55.4	19.7	53.9	55.7	37.2	44.4	25.9	69.7	63.2	45.4	50.9	51.0
$K = 10$	FSRW^[11]	30.0	62.4	43.2	60.1	39.7	47.1	43.1	13.5	41.6	58.4	39.5	39.2	20.5	51.7	55.9	42.3	36.2	41.3
	MetaDet^[22]	52.4	56.1	51.5	55.8	41.3	51.4	52.8	3.3	52.1	69.8	45.2	44.6	13.6	71.3	58.6	50.4	47.4	48.3
	MPSR^[15]	48.2	72.8	68.2	71.2	48.3	61.7	51.6	16.4	54.8	65.4	47.1	47.1	24.8	55.3	67.1	50.5	47.2	49.6
	Ours	59.4	67.8	69.3	69.8	44.6	62.2	59.4	19.6	58.9	66.6	59.6	52.8	34.2	73.1	69.5	52.8	55.9	57.1

表2 VOC07测试集上的 AP 和 mAP

Table 2 AP and mAP on VOC07 test suite %

Shot	Methods	Novel-class split 1						Novel-class split 2						Novel-class split 3
Shot	Methods	bird	bus	cow	mobike	sofa	mean	aero	bottle	cow	horse	sofa	mean	boat	cat	mobike	sheep	sofa	mean
$K = 1$	FSRW^[11]	13.2	15.5	30.7	35.2	7.9	20.5	11.7	9.1	15.2	23.9	18.4	15.7	10.5	44.0	18.4	18.3	5.2	19.3
	MetaDet^[22]	6.2	32.5	15.4	35.6	0.7	18.1	23.5	0.9	23.8	4.8	0.9	10.8	4.6	31.4	11.5	11.8	0.2	11.9
	MPSR^[15]	33.8	41.9	57.2	53.0	21.4	41.5	21.4	9.5	35.9	30.6	25.3	24.5	14.3	47.8	35.4	32.4	22.4	30.5
	Ours	48.6	39.4	44.6	53.5	21.5	41.5	31.8	7.6	27.6	39.4	38.7	29.0	19.7	55.9	17.2	18.2	43.9	31.0
$K = 2$	FSRW^[11]	21.3	12.4	16.7	17.2	9.3	15.4	28.5	0.9	27.3	0.0	19.2	15.2	5.2	46.5	18.9	26.7	12.0	21.9
	MetaDet^[22]	35.2	47.5	54.8	32.2	0.4	34.0	12.3	1.0	44.1	52.9	0.8	22.2	16.5	25.8	19.5	0.9	18.5	16.2
	MPSR^[15]	38.5	16.4	56.3	57.3	32.4	40.2	36.4	9.1	46.8	27.5	34.2	30.8	18.7	45.3	59.7	49.2	32.8	41.1
	Ours	48.3	55.7	52.0	59.5	29.8	49.1	36.5	9.7	51.2	55.7	17.9	34.2	15.9	69.7	59.9	41.4	37.2	44.8
$K = 3$	FSRW^[11]	26.2	19.4	20.5	20.5	27.0	27.7	29.4	4.9	34.5	6.7	37.6	22.6	11.3	39.5	20.7	23.1	33.6	25.6
	MetaDet^[22]	32.4	46.1	53.85	38.4	10.7	36.3	26.8	0.3	50.6	53.8	18.9	30.1	13.6	39.3	32.8	38.4	10.5	26.9
	MPSR^[15]	35.1	60.7	61.4	61.3	43.6	52.5	47.6	9.0	47.3	45.2	40.5	37.9	14.9	62.8	57.9	37.5	42.1	43.0
	Ours	59.4	67.8	69.3	64.9	45.2	61.3	45.4	15.6	47.6	46.6	38.9	38.8	24.2	65.1	55.9	52.8	36.8	47.0
$K = 5$	FSRW^[11]	31.6	22.5	39.6	40.0	37.1	34.2	33.1	9.6	38.4	25.9	44.0	30.2	14.5	57.6	50.2	38.6	41.5	40.5
	MetaDet^[22]	35.2	47.5	54.8	55.1	34.8	45.5	28.3	1.7	50.6	54.9	38.4	34.8	16.5	45.8	53.1	40.3	48.5	40.8
	MPSR^[15]	37.9	65.5	55.1	68.6	47.5	54.9	47.6	10.7	45.1	45.8	47.6	39.4	20.7	56.8	68.1	48.7	45.6	48.0
	Ours	58.3	65.7	43.0	66.7	41.5	55.0	55.4	19.7	53.9	55.7	37.2	44.4	25.9	69.7	63.2	45.4	50.9	51.0
$K = 10$	FSRW^[11]	30.0	62.4	43.2	60.1	39.7	47.1	43.1	13.5	41.6	58.4	39.5	39.2	20.5	51.7	55.9	42.3	36.2	41.3
	MetaDet^[22]	52.4	56.1	51.5	55.8	41.3	51.4	52.8	3.3	52.1	69.8	45.2	44.6	13.6	71.3	58.6	50.4	47.4	48.3
	MPSR^[15]	48.2	72.8	68.2	71.2	48.3	61.7	51.6	16.4	54.8	65.4	47.1	47.1	24.8	55.3	67.1	50.5	47.2	49.6
	Ours	59.4	67.8	69.3	69.8	44.6	62.2	59.4	19.6	58.9	66.6	59.6	52.8	34.2	73.1	69.5	52.8	55.9	57.1

图6 在不同 K下的平均精度

Fig.6 Average accuracy under different K

表3 COCO2014测试集上的 AP 和 mAP

Table 3 AP and mAP on COCO2014 test suite %

Shot	Methods	$A P$	$A P 50$	$A P 75$	$A P S$	$A P M$	$A P L$
$K = 10$	FSRW^[11]	5.5	12.4	4.8	0.7	3.3	10.7
	MetaDet^[22]	7.4	13.9	6.2	1.3	4.0	12.5
	FSDetView^[24]	12.3	27.6	9.8	2.4	7.6	13.9
	MPSR^[15]	9.4	17.8	9.8	3.4	9.0	15.8
	DCNet^[27]	12.5	23.8	11.6	4.5	13.7	21.1
	Meta R-CNN^[3]	8.4	19.2	6.4	2.2	7.8	14.3
	Ours	15.9	28.4	13.5	3.9	15.4	24.6
$K = 30$	FSRW^[11]	9.3	19.5	7.4	0.9	4.8	16.5
	MetaDet^[22]	11.5	21.7	8.4	1.3	6.5	17.8
	FSDetView^[24]	14.2	30.9	12.4	3.7	15.9	23.4
	MPSR^[15]	14.0	28.7	14.2	4.1	12.5	22.8
	DCNet^[27]	18.9	32.7	17.0	6.9	16.4	26.7
	Meta R-CNN^[3]	12.3	25.6	11.1	2.5	11.5	19.1
	Ours	17.8	32.9	16.3	4.9	17.2	26.8

表3 COCO2014测试集上的 AP 和 mAP

Table 3 AP and mAP on COCO2014 test suite %

Shot	Methods	$A P$	$A P 50$	$A P 75$	$A P S$	$A P M$	$A P L$
$K = 10$	FSRW^[11]	5.5	12.4	4.8	0.7	3.3	10.7
	MetaDet^[22]	7.4	13.9	6.2	1.3	4.0	12.5
	FSDetView^[24]	12.3	27.6	9.8	2.4	7.6	13.9
	MPSR^[15]	9.4	17.8	9.8	3.4	9.0	15.8
	DCNet^[27]	12.5	23.8	11.6	4.5	13.7	21.1
	Meta R-CNN^[3]	8.4	19.2	6.4	2.2	7.8	14.3
	Ours	15.9	28.4	13.5	3.9	15.4	24.6
$K = 30$	FSRW^[11]	9.3	19.5	7.4	0.9	4.8	16.5
	MetaDet^[22]	11.5	21.7	8.4	1.3	6.5	17.8
	FSDetView^[24]	14.2	30.9	12.4	3.7	15.9	23.4
	MPSR^[15]	14.0	28.7	14.2	4.1	12.5	22.8
	DCNet^[27]	18.9	32.7	17.0	6.9	16.4	26.7
	Meta R-CNN^[3]	12.3	25.6	11.1	2.5	11.5	19.1
	Ours	17.8	32.9	16.3	4.9	17.2	26.8

图7 混合平均精度对比

Fig.7 Comparison of mixed average accuracy

表4 不同模块在VOC07测试集上的low-shot检测mAP

Table 4 Low-shot detection mAP of different modules on VOC07 test suite %

Components	VOC07 split 1
Components	$K = 1$	$K = 2$	$K = 3$	$K = 5$	$K = 10$
Base	20.0	25.4	38.4	44.8	50.5
Base+FSE	22.3	28.2	40.9	48.3	52.1
Base+FSS	23.1	29.1	41.7	49.6	53.4
Base+FSE +FSS	23.7	29.7	42.8	50.2	54.3
Base+MFF	32.6	34.9	46.0	47.4	52.8
Base+MFF+FS	40.9	45.8	48.9	54.7	57.9
Base+MFF+FS+OL	42.3	49.3	52.7	58.0	64.4
Base+FSE+FSS+MFF+FS+OL	43.7	50.4	53.6	59.8	66.4

表4 不同模块在VOC07测试集上的low-shot检测mAP

Table 4 Low-shot detection mAP of different modules on VOC07 test suite %

Components	VOC07 split 1
Components	$K = 1$	$K = 2$	$K = 3$	$K = 5$	$K = 10$
Base	20.0	25.4	38.4	44.8	50.5
Base+FSE	22.3	28.2	40.9	48.3	52.1
Base+FSS	23.1	29.1	41.7	49.6	53.4
Base+FSE +FSS	23.7	29.7	42.8	50.2	54.3
Base+MFF	32.6	34.9	46.0	47.4	52.8
Base+MFF+FS	40.9	45.8	48.9	54.7	57.9
Base+MFF+FS+OL	42.3	49.3	52.7	58.0	64.4
Base+FSE+FSS+MFF+FS+OL	43.7	50.4	53.6	59.8	66.4

图8 部分测试结果对比

Fig.8 Comparison of partial test results

图9 正交损失不同参数的mAP对比图

Fig.9 mAP of different parameters of orthogonal loss

参考文献 27

[1]	GIRSHICK R B, DONAHUE J, DARRELL T, et al. Rich fea-ture hierarchies for accurate object detection and semantic segmentation[C]// Proceedings of the 2014 IEEE Conference on Computer Vision and Pattern Recognition, Columbus, Jun 23-28, 2014. Washington: IEEE Computer Society, 2014: 580-587.
[2]	GIRSHICK R B. Fast R-CNN[C]// Proceedings of the 14th IEEE International Conference on Computer Vision, Santiago, Dec 7-13, 2015. Washington: IEEE Computer Society, 2015: 1440-1448.
[3]	YAN X P, CHEN Z L, XU A N, et al. Meta R-CNN: towards general solver for instance-level low-shot learning[C]// Pro-ceedings of the 2019 IEEE/CVF International Conference on Computer Vision, Seoul, Oct 27-Nov 2, 2019. Piscataway: IEEE, 2019: 9577-9586.
[4]	REDMON J, DIVVALA S K, GIRSHICK R B, et al. You only look once: unified, real-time object detection[C]// Procee-dings of the 2016 IEEE Conference on Computer Vision and Pattern Recognition, Las Vegas, Jun 27-30, 2016. Was-hington: IEEE Computer Society, 2016: 779-788.
[5]	祝钧桃, 姚光乐, 张葛祥, 等. 深度神经网络的小样本学习综述[J]. 计算机工程与应用, 2021, 57(7): 22-33.
	ZHU J T, YAO G L, ZHANG G X, et al. Survey of few shot learning of deep neural network[J]. Computer Engineering and Applications, 2021, 57(7): 22-33.
[6]	许德刚, 王露, 李凡. 深度学习的典型目标检测算法研究综述[J]. 计算机工程与应用, 2021, 57(8): 10-25.
	XU D G, WANG L, LI F. Review of typical object detection algorithms for deep learning[J]. Computer Engineering and Applications, 2021, 57(8): 10-25.
[7]	CHEN H, WANG Y L, WANG G Y, et al. LSTD: a low-shot transfer detector for object detection[C]// Proceedings of the 32nd AAAI Conference on A.pngicial Intelligence, the 30th Innovative Applications of A.pngicial Intelligence, and the 8th AAAI Symposium on Educational Advances in Arti-ficial Intelligence, New Orleans, Feb 2-7, 2018. Menlo Park: AAAI, 2018: 2836-2843.
[8]	REN S Q, HE K M, GIRSHICK R B, et al. Faster R-CNN:towards real-time object detection with region proposal net-works[C]// Advances in Neural Information Processing Sys-tems 28, Montreal, Dec 7-12, 2015. Red Hook: Curran Asso-ciates, 2015: 91-99.
[9]	LIU W, ANGUELOV D, ERHAN D, et al. SSD: single shot multibox detector[C]// LNCS 9905: Proceedings of the 14th European Conference on Computer Vision, Amsterdam, Oct 11-14, 2016. Cham: Springer, 2016: 21-37.
[10]	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 Conference on Computer Vision and Pattern Recognition, Long Beach, Jun 16-20, 2019. Piscataway: IEEE, 2019: 5197-5206.
[11]	KANG B Y, LIU Z, WANG X, et al. Few-shot object detec-tion via feature reweighting[C]// Proceedings of the 2019 IEEE/CVF International Conference on Computer Vision, Seoul, Oct 27-Nov 2, 2019. Piscataway: IEEE, 2019: 8419-8428.
[12]	FAN Q, ZHUO W, TANG C K, et al. Few-shot object detec-tion with attention-RPN and multi-relation detector[C]// Pro-ceedings of the 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, Seattle, Jun 13-19, 2020. Pis-cataway: IEEE, 2020: 4012-4021.
[13]	VINYALS O, BLUNDELL C, LILLICRAP T, et al. Matching networks for one shot learning[C]// Advances in Neural In-formation Processing Systems 29, Barcelona, Dec 5-10, 2016. Red Hook: Curran Associates, 2016: 3630-3638.
[14]	LI Y T, CHENG Y, LIU L, et al. Low-shot object detection via classification refinement[J]. arXiv: 2005. 02641, 2020.
[15]	WU J X, LIU S T, HUANG D, et al. Multi-scale positive sample refinement for few-shot object detection[C]// LNCS 12361: Proceedings of the 16th European Conference on Computer Vision, Glasgow, Aug 23-28, 2020. Cham: Springer, 2020: 456-472.
[16]	韩子硕, 王春平, 付强, 等. 基于超密集特征金字塔网络的SAR图像舰船检测[J]. 系统工程与电子技术, 2020, 42(10): 2214-2222.
	HAN Z S, WANG C P, FU Q, et al. Ship detection in SAR images based on super dense feature pyramid networks[J]. Systems Engineering and Electronics, 2020, 42(10): 2214-2222.
[17]	YANG Z, WANG Y L, CHEN X Y, et al. Context-Transformer: tackling object confusion for few-shot detection[C]// Procee-dings of the 34th AAAI Conference on A.pngicial Intelli-gence, the 32nd Innovative Applications of A.pngicial Intelli-gence Conference, the 10th AAAI Symposium on Educatio-nal Advances in A.pngicial Intelligence, New York, Feb 7-12, 2020. Menlo Park: AAAI, 2020: 12653-12660.
[18]	KIM G, JUNG H G, LEE S W. Few-shot object detection via knowledge transfer[C]// Proceedings of the 2020 IEEE International Conference on Systems, Man, and Cybernetics, Toronto, Oct 11-14, 2020. Piscataway: IEEE, 2020: 3564-3569.
[19]	HE K M, GKIOXARI G, DOLLÁR P, et al. Mask R-CNN[C]// Proceedings of the 2017 IEEE International Conference on Computer Vision, Venice, Oct 22-29, 2017. Washington: IEEE Computer Society, 2017: 2961-2969.
[20]	LIN T Y, MAIRE M, BELONGIE S J, et al. Microsoft COCO: common objects in context[C]// LNCS 8693: Pro-ceedings of the 13th European Conference on Computer Vi-sion, Zurich, Sep 6-12, 2014. Cham: Springer. 2014: 740-755.
[21]	HE K, ZHANG X, REN S, et al. Deep residual learing for image recognition[J]. arXiv:1512. 03385, 2015.
[22]	WANG Y X, RAMANAN D, HEBERT M. Meta-learning to detect rare objects[C]// Proceedings of the 2019 IEEE/CVF International Conference on Computer Vision, Seoul, Oct 27-Nov 2, 2019. Piscataway: IEEE, 2019: 9924-9933.
[23]	WANG X, HUANG T E, DARRELL T, et al. Frustratingly simple few-shot object detection[J]. arXiv: 2003. 06957, 2020.
[24]	XIAO Y, MARLET R. Few-shot object detection and view-point estimation for objects in the wild[C]// LNCS 12362: Proceedings of the 16th European Conference on Computer Vision, Glasgow, Aug 23-28, 2020. Cham: Springer, 2020: 192-210.
[25]	LIU L Y, MA B, ZHANG Y L, et al. AFD-Net: adaptive fully-dual network for few-shot object detection[C]// Procee-dings of the 29th ACM International Conference on Multi-media, Oct 20-24, 2021. New York: ACM, 2021: 2549-2557.
[26]	LI B H, YANG B Y, LIU C, et al. Beyond max-margin: class margin equilibrium for few-shot object detection[C]// Pro-ceedings of the 2021 IEEE Conference on Computer Vision and Pattern Recognition, Jun 19-25, 2021. Piscataway: IEEE, 2021: 7363-7372.
[27]	HU H Z, BAI S, LI A X, et al. Dense relation distillation with context-aware aggregation for few-shot object detec-tion[C]// Proceedings of the 2021 IEEE Conference on Com-puter Vision and Pattern Recognition, Jun 19-25, 2021. Pis-cataway: IEEE, 2021: 10185-10194.