基于解耦的小样本目标检测方法研究

doi:10.12305/j.issn.1001-506X.2024.09.06

摘要/Abstract

摘要：

针对小样本激化的目标检测耦合问题, 以高价值空中目标为研究对象, 提出一种基于解耦的小样本目标检测算法。首先, 在区域候选网络中引入梯度调整层, 强化区域候选网络, 缓和任务耦合问题。其次, 将目标检测头拆解成分类和回归两个分支, 在前端添加无参平均注意力模块, 缓和特征耦合问题。所提算法可以提高小样本目标检测性能, 增强对新类的检测能力。实验结果表明, 所提算法在1、2、3、5、10样本实验中均表现最佳, 平均精度分别达到32.5%、35.6%、39.6%、41.2%和57.4%。相较于两阶段微调方法, 所提算法检测性能大幅度提高, 能够解决在小样本激化的耦合矛盾下网络检测能力下降的问题, 提升对小样本高价值空中目标的检测精度

关键词: 小样本目标检测, 空中目标, 耦合问题, 深度学习

Abstract:

Aiming at the coupling problem of target detection with few shot intensification, a few shot target detection algorithm based on decoupling is proposed focusing on high-value air targets as the research object. Firstly, a gradient adjustment layer is introduced into the regional candidate network to strengthen the regional candidate network and alleviate the task coupling problem. Secondly, the target detection head is disassembled into two branches, classification and regression, and a parameter-free average attention module is added at the front end to alleviate the feature coupling problem. The proposed algorithm improves the detection performance of few shot target detection and enhances the detection ability of new classes. The experimental results show that the proposed algorithm performs best in the 1, 2, 3, 5 and 10 shot experiments, with average accuracy of 32.5%, 35.6%, 39.6%, 41.2% and 57.4%, respectively. Compared with the two-stage fine-tuning method, the detection performance of the proposed algorithm is greatly improved, which solves the problem of the decline of network detection ability under the coupling contradiction of few shot intensification, and improves the detection accuracy of few shot high-value air targets.

Key words: few shot target detection, air target, coupling issue, deep learning

中图分类号:

TP391.4

蔡伟, 王鑫, 蒋昕昊, 杨志勇, 陈栋. 基于解耦的小样本目标检测方法研究[J]. 系统工程与电子技术, 2024, 46(9): 2941-2950.

Wei CAI, Xin WANG, Xinhao JIANG, Zhiyong YANG, Dong CHEN. Research on few shot target detection method based on decoupling[J]. Systems Engineering and Electronics, 2024, 46(9): 2941-2950.

图/表 14

图1

图2

图3

图4

图5

图6

图7

表1

表2

图8

表3

表4

表5

图9

参考文献 41

1	杜芸彦, 李鸿, 杨锦辉, 等. 基于负边距损失的小样本目标检测[J]. 计算机应用, 2022, 42 (11): 3617- 3624.
	DU Y Y , LI H , YANG J H , et al. Few shot target detection based on negative margin loss[J]. Computer Applications, 2022, 42 (11): 3617- 3624.
2	KAUL P, XIE W D, ZISSERMAN A. Label, verify, correct: a simple few shot object detection method[C]//Proc. of the IEEE Conference on Computer Vision and Pattern Recognition, 2022: 14237-14247.
3	YIN L, PEREZ-RUA J M, LIANG K J. Sylph: a hypernetwork framework for incremental few-shot object detection[C]//Proc. of the IEEE Conference on Computer Vision and Pattern Recognition, 2022: 9035-9045.
4	ZHANG S, WANG L, MURRAY N, et al. Kernelized few-shot object detection with efficient integral aggregation[C]//Proc. of the IEEE Conference on Computer Vision and Pattern Recognition, 2022: 19207-19216.
5	CHEN T I , LIU Y C , SU H T , et al. Dual-awareness attention for few-shot object detection[J]. IEEE Trans.on Multimedia, 2023, 25 (1): 291- 301.
6	FAN Z B, MA Y C, LI Z M, et al. Generalizedfew-shot object detection without forgetting[C]//Proc. of the IEEE Conference on Computer Vision and Pattern Recognition, 2021: 4527-4536.
7	徐鹏帮, 桑基韬, 路冬媛. 类别语义相似性监督的小样本图像识别[J]. 中国图象图形学报, 2021, 26 (7): 1594- 1603.
	XU P B , SANG J T , LU D Y . Few shot image recognition for category semantic similarity monitoring[J]. Chinese Journal of Image Graphics, 2021, 26 (7): 1594- 1603.
8	张振伟, 郝建国, 黄健, 等. 小样本图像目标检测研究综述[J]. 计算机工程与应用, 2022, 58 (5): 1- 11.
	ZHANG Z W , HAO J G , HUANG J , et al. Summary of research on target detection in small sample images[J]. Computer Engineering and Applications, 2022, 58 (5): 1- 11.
9	SUN B, LI B H, CAI S C, et al. Few-shot object detection via contrastive proposal encoding[C]//Proc. of the IEEE Conference on Computer Vision and Pattern Recognition, 2021: 7352-7362.
10	WU A M, HAN Y H, ZHU L C, et al. Universal-prototype enhancing for few-shot object detection[C]//Proc. of the International Conference on Computer Vision, 2021: 9567-9576.
11	LI X, ZHANG L, CHEN Y P, et al. One-shot object detection without fine-tuning[EB/OL]. [2023-03-16]. http://arxiv.org/abs/2005.03819.
12	HSIEH T I, LO Y C, CHEN H T, et al. One-shot object detection with co-attentionand co-excitation[EB/OL]. [2023-03-16]. http://arxiv.org/abs/1911.12529.
13	WANG X, HUANG T E, DARRELL T, et al. Frustratingly simple few shot object detection[C]//Proc. of the 37th International Conference on Machine Learning, 2020: 9919-9928.
14	GIRSHICK R, DONAHUE J, DARRELL T, et al. Rich feature hierarchies for accurate object detection and semantic segmentation[C]//Proc. of the IEEE Conference on Computer Vision and Pattern Recognition, 2014: 580-587.
15	GIRSHICK R. Fast R-CNN[C]//Proc. of the IEEE International Conference on Computer Vision, 2015: 1440-1448.
16	REN S , HE K , GIRSHICK R , et al. Faster RCNN: towards real-time object detection with region proposal networks[J]. IEEE Trans.on Pattern Analysis & Machine Intelligence, 2017, 39 (6): 1137- 1149.
17	范加利, 田少兵, 黄葵, 等. 基于Faster R-CNN的航母舰面多尺度目标检测算法[J]. 系统工程与电子技术, 2022, 44 (1): 40- 46. doi: 10.12305/j.issn.1001-506X.2022.01.06
	FAN J L , TIAN S B , HUANG K , et al. Multiscale target detection algorithm for aircraft carrier surface based on Fast R-CNN[J]. Systems Engineering and Electronics, 2022, 44 (1): 40- 46. doi: 10.12305/j.issn.1001-506X.2022.01.06
18	张大恒, 张英俊, 张闯. 基Faster R-CNN的电子海图和雷达图像的数据融合[J]. 系统工程与电子技术, 2020, 42 (6): 1267- 1273. doi: 10.3969/j.issn.1001-506X.2020.06.09
	ZHANG D H , ZHANG Y J , ZHANG C . Data fusion of electronic charts and radar images based on fast R-CNN[J]. Systems Engineering and Electronics, 2020, 42 (6): 1267- 1273. doi: 10.3969/j.issn.1001-506X.2020.06.09
19	JIANG X H , CAI W , YANG Z Y , et al. A lightweight multiscale infrared aerocraft recognition algorithm[J]. Arabian Journal for Science and Engineering, 2022, 47 (2): 2289- 2303. doi: 10.1007/s13369-021-06181-7
20	JIANG X H , CAI W , DING Y , et al. Camouflaged object detection based on ternary cascade perception[J]. Remote Sensing, 2023, 15 (5): 1188. doi: 10.3390/rs15051188
21	REDMIN J, DIVALAS, GIRSICK R, et al. You can only watch it once: unified real time object detection[C]//Proc. of the IEEE Conference on Computer Vision and Pattern Recognition, 2016: 777-7888.
22	REDMON J, FARHADI A. YOLO9000: better, faster, stronger[C]//Proc. of the IEEE Conference on Computer Vision and Pattern Recognition, 2017: 7263-7271.
23	REDMON J, FARHADI A. YOLOv3: an incremental improve- ment[EB/OL]. [2023-03-16]. http://arxiv.org/abs/1804.02767.
24	ALEXEY B, WANG C Y, HONG Y. Yolov4: optimal speed and accuracy of object detection[EB/OL]. [2023-03-16]. http://arxiv.org/abs/2004.10934.
25	LIU W, ANGUELOV D, ERHAN D, et al. SSD: single shot multibox detector[C]//Proc. of the European Conference on Computer Vision, 2016: 21-37.
26	FU C Y, LIU W, RANGA A, et al. DSSD: deconvolutional single shot detector[EB/OL]. [2023-03-16]. http://arxiv.org/abs/1701.06659.
27	XIE S N, GIRSHICK R, DOLLAR P, et al. Aggregated resi-dual transformations for deep neural networks[C]//Proc. of the IEEE International Conference on Computer Vision and Pattern Recognition, 2017: 1492-1500.
28	FINN C, ABBEEL P, LEVINE S. Model-agnostic meta-learning for fast adaptation of deep network[EB/OL]. [2023-03-16]. http://arxiv.org/abs/1703.03400.
29	JAMAL M A, QI G J. Task agnostic meta-learning for few-shot learning[C]//Proc. of the IEEE Conference on Computer Vision and Pattern Recognition, 2019: 11719-11727.
30	KARLINSKY L, SHTOK J, HARARY S, et al. RepMet: representative based metric learning for classification and few-shot object detection[C]//Proc. of the IEEE Conference on Computer Vision and Pattern Recognition, 2019: 5197-5206.
31	WANG P, LIU L U, SHEN C H, et al. Multiple-attention networks for one learning[C]//Proc. of the IEEE Conference on Computer Vision and Pattern Recognition, 2017: 2721-2729.
32	TIAN Y L, WANG Y L, KRISHNAN D, et al. Rethinking few-shot image classification: agood embedding is all you need?[C]//Proc. of the European Conference on Computer Vision, 2020: 266-282.
33	SUN Q R, LIU Y Y, CHUA T S, et al. Metatransfer learning for few-shot learning[C]//Proc. of the IEEE Conference on Computer Vision and Pattern Recognition, 2019: 403-412.
34	LIU Y Y, SUN Q R, LIU A A, et al. LCC: learning to customize and combine neural networks for few-shot Learning[EB/OL]. [2023-03-16]. http://arxiv.org/abs/1904.08479.
35	KANG B Y, LIU Z, WANG X, et al. Few shot object detection via feature reweighting[C]//Proc. of the IEEE International Conference on Computer Vision, 2019: 8420-8429.
36	XIAO Y, LETEPIT V, MARLET R. Few-shot object detection and viewpoint estimation for objectsin the wild[C]//Proc. of the European Conference on Computer Vision, 2020: 192-210.
37	YAN X P, CHEN Z L, XU A N, et al. Meta R-CNN: towards general solver for instancelevel low-shot learning[C]//Proc. of the IEEE International Conference on Computer Vision, 2019: 9577-9586.
38	FAN Q, ZHUO W, TANG C K, et al. Few shot object detection with attention RPN and multi-relation detector[C]//Proc. of the IEEE Conference on Computer Vision and Pattern Recogni- tion, 2020: 4013-4022.
39	WU J X, LIU S T, HUANG D, et al. Multi scale positive sample refinement for fewshot object detection[C]//Proc. of the European Conference on Computer Vision, 2020: 456-472.
40	KHANDELWAL S, GOYAL R, SIGAl L. Unified knowledge transfer for any shot object detection and segmentation[C]//Proc. of the IEEE Conference on Computer Vision and Pattern Recognition, 2021: 5951-5961.
41	ZHANG W L, WANG Y X, FORSYTH D A. Cooperating RPN's improve few-shot object detection[EB/OL]. [2023-03-16]. http://arxiv.org/abs/2011.10142.

方法	样本量
方法	1	2	3	5	10
TFA/FCL	25.6	29.4	27.4	37.1	44.2
TFA/COS	17.2	23.5	23.1	23.4	29.2
Attention RPN	23.8	29.2	29.1	39.4	54.3
本文	32.5	35.6	39.6	41.2	57.4

方法	目标种类						mAP
方法	F35	J20	Su57	MQ9	RQ4	B2	mAP
TFA/FCL	12.3	55.5	35.6	60.9	43	55.8	44.2
TFA/COS	35.9	45.6	24.3	37.6	16	7.8	29.2
Attention RPN	9.8	61.5	54.2	62.5	65.5	72.5	54.3
本文	19.8	63.3	74.4	61	68.4	62.7	57.4

基线模型	筛选原则	解冻	PFAAM	DBN	样本量
基线模型	筛选原则	解冻	PFAAM	DBN	1	2	3	5	10
√	-	-	-	-	25.6	29.4	27.4	37.1	44.2
√	√	-	-	-	27.5	32.6	35.3	39.4	44.6
√	√	√	-	-	28.1	33.6	36.4	39.6	50.9
√	√	√	√	-	28.5	33.7	36.6	40.2	52.7
√	√	√	√	√	30.5	34.2	38.4	40.7	54.2

调整因子	样本量
调整因子	1	2	3	5	10
0.25	31.8	34.7	39.2	41	55.7
0.5	32.5	35.6	39.6	41.2	57.4
0.75	31.5	34.8	38.8	40.8	56.6
1	30.5	34.2	38.4	40.7	54.2

方法	样本量
方法	1	2	3	5	10
TFA/FCL	36.8	39.1	43.6	55.7	57
TFA/COS	39.8	36.1	44.7	55.7	56
Attention RPN	35	36	39.1	51.7	55.7
本文	40.1	45.3	51.1	58.1	60