基于多阶特征聚合的小样本虫害检测

doi:10.19788/j.issn.2096-6369.000141

摘要/Abstract

摘要：

在实际农田场景中，部分虫害因难以采集而导致样本数量稀少，传统检测方法难以在小样本条件下获得可靠的检测与分类效果。本研究旨在针对小样本虫害检测问题，提升模型的泛化能力与检测性能。构建了包含13类虫害的AgriPest-13数据集，其中包括11类常见虫害与2类少样本虫害（分别仅含10张与30张图像）。在此基础上，提出改进的小样本检测方法FSD-MOGA（Few-Shot Detection with Multi-Order Gated Aggregation）。该方法以小样本检测模型MPSR为基础，引入多阶门控聚合模块（MOGA）以增强复杂背景下的特征表征能力，并加入中心损失（Center Loss）以优化分类特征分布。实验表明，FSD-MOGA在小样本场景下表现优异，尤其在仅有极少样本的类别中表现更为稳定和准确：10-shot实验中mAP达到37.0%，较基线模型MPSR提升9.4%，AP50、AP75分别达到57.8%和29.9%；30-shot实验中mAP进一步提升至44.0%，较MPSR提高2.6%，AP50、AP75升至73.6%和45.5%。FSD-MOGA能有效缓解虫害样本稀缺带来的检测困难，为实际农田场景下的小样本虫害智能检测提供了一种可行且有效的技术方案。

关键词: 小样本检测, 虫害识别, 特征聚合, Center Loss, 智慧农业

Abstract:

In real farmland scenarios, some pest species are difficult to collect, resulting in very limited sample sizes. Traditional detection methods struggle to achieve reliable detection and classification performance under such few-shot conditions. This study aims to address the problem of few-shot pest detection by improving the model’s generalization ability and detection performance. A dataset named AgriPest-13 containing 13 pest categories was constructed, including 11 common pest species and 2 few-shot pests with only 10 and 30 images respectively. Based on this dataset, an improved few-shot detection method, FSD-MOGA (Few-Shot Detection with Multi-Order Gated Aggregation), is proposed. The method is developed upon the classic few-shot detection model MPSR, integrating a Multi-Order Gated Aggregation (MOGA) module to enhance feature representation under complex backgrounds and incorporating Center Loss to optimize the distribution of classification features. Compared with the original MPSR model, FSD-MOGA achieves significant performance improvements in few-shot pest detection tasks, particularly demonstrating more stable and accurate detection for categories with extremely limited samples. FSD-MOGA effectively alleviates the detection challenges caused by scarce pest samples and provides a feasible and efficient solution for intelligent pest detection in real agricultural environments with limited data.

Key words: few-shot detection, pest recognition, feature aggregation, Center Loss, smart agriculture

曹佩佩, 李贺, 任保见. 基于多阶特征聚合的小样本虫害检测[J]. 农业大数据学报, 2026, 8(3): 296-304.

CAO PeiPei, LI He, REN BaoJian. Few-Shot Pest Detection via Multi-Order Feature Aggregation[J]. Journal of Agricultural Big Data, 2026, 8(3): 296-304.

图/表 9

表1

图1

图2

图3

图4

表2

表3

图5

图6

参考文献 23

[1]	Strange R N, Scott P R. Plant disease: a threat to global food security. Annual Review of Phytopathology, 2005, 43: 83-116. pmid: 16078878
[2]	Ahmed S, Marwat S N K, Brahim G B, et al. IoT based intelligent pest management system for precision agriculture. Scientific Reports, 2024, 14(1): 31917. doi: 10.1038/s41598-024-83012-3
[3]	Weeks P J D, O’Neill M A, Gaston K J, et al. Automating insect identification: exploring the limitations of a prototype system. Journal of Applied Entomology, 1999, 123(1): 1-8. doi: 10.1046/j.1439-0418.1999.00307.x
[4]	Gaston K J, O’Neill M A. Automated species identification: why not?. Philosophical Transactions of the Royal Society of London. Series B, Biological Sciences, 2004, 359(1444): 655-667.
[5]	Hayashi M, Tamai K, Owashi Y, et al. Automated machine learning for identification of pest aphid species (Hemiptera: Aphididae). Applied Entomology and Zoology, 2019, 54(4): 487-490. doi: 10.1007/s13355-019-00642-0
[6]	Li W, Zheng T, Yang Z, et al. Classification and detection of insects from field images using deep learning for smart pest management: A systematic review. Ecological Informatics, 2021, 66: 101460. doi: 10.1016/j.ecoinf.2021.101460
[7]	Li W, Wang D, Li M, et al. Field detection of tiny pests from sticky trap images using deep learning in agricultural greenhouse. Computers and Electronics in Agriculture, 2021, 183: 106048. doi: 10.1016/j.compag.2021.106048
[8]	Preti M, Verheggen F, Angeli S. Insect pest monitoring with camera- equipped traps: strengths and limitations. Journal of Pest Science, 2021, 94(2): 203-217. doi: 10.1007/s10340-020-01309-4
[9]	彭红星, 黄博, 邵园园, 等. 自然环境下多类水果采摘目标识别的通用改进SSD模型. 农业工程学报, 2018, 34(16): 155-162.
	Peng H X, Huang B, Shao Y Y, et al. General improved SSD model for picking object recognition of multiple fruits in natural environment. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2018, 34(16): 155-162.
[10]	Ge Y, Lin S, Zhang Y, et al. Tracking and counting of tomato at different growth period using an improving YOLO-deepsort network for inspection robot. Machines, 2022, 10(6): 489. doi: 10.3390/machines10060489
[11]	宫妍, 翟俊杰, 王凯, 等. 基于改进YOLOv7的玉米作物害虫检测研究. 计算机测量与控制, 2024, 32(9): 58-65.
	Gong Y, Zhai J J, Wang K, et al. Study on corn crop pest detection based on improved YOLOv7. Computer Measurement & Control, 2024, 32(9): 58-65.
[12]	Teng Y, Zhang J, Dong S F, et al. MSR-RCNN: a multi-class crop pest detection network based on a multi-scale super-resolution feature enhancement module. Frontiers in Plant Science, 2022, 13: 810546. doi: 10.3389/fpls.2022.810546
[13]	Qi F, Chen G M, Liu J Y, et al. End-to-end pest detection on an improved deformable DETR with multihead criss cross attention. Ecological Informatics, 2022, 72: 101902. doi: 10.1016/j.ecoinf.2022.101902
[14]	Bai Y Z, Hou F J, Fan X Y, et al. A lightweight pest detection model for drones based on transformer and super-resolution sampling techniques. Agriculture, 2023, 13(9): 1812. doi: 10.3390/agriculture13091812
[15]	Liu S T, Fu S H, Hu A R, et al. Research on insect pest identification in rice canopy based on GA-mask R-CNN. Agronomy, 2023, 13(8): 2155. doi: 10.3390/agronomy13082155
[16]	Zhou Y, Yan H, Ding K, et al. Few-shot image classification of crop diseases based on vision-language models. Sensors (Basel), 2024, 24(18): 6109. doi: 10.3390/s24186109
[17]	He S, Jin B, Sun X, et al. Few-shot object detection for pest insects via features aggregation and contrastive learning. Frontiers in Plant Science, 2025, 16:1522510. DOI: 10.3389/fpls.2025.1522510
[18]	李开敬, 许燕, 周建平, 等. 基于Faster R-CNN和数据增强的棉田苗期杂草识别方法. 新疆大学学报(自然科学版)(中英文), 2021, 38(4): 450-456.
	Li K J, Xu Y, Zhou J P, et al. Cotton field seedling weed identification method based on faster R-CNN and data enhancement. Journal of Xinjiang University (Natural Science Edition in Chinese and English), 2021, 38(4): 450-456.
[19]	Redmon J, Farhadi A.YOLOv3: An Incremental Improvement[PP/OL]. 2018, arXiv.1804.02767.
[20]	Zhang H, Li F, Liu S L, et al. DINO: DETR With improved DeNoising anchor boxes for end-to-end object detection[PP/OL]. 2022, arXiv./2203.03605.
[21]	Huisman M, Van Rijn J N, Plaat A. A survey of deep Meta-learning. Artificial Intelligence Review, 2021, 54(6): 4483-4541. doi: 10.1007/s10462-021-10004-4
[22]	Köhler M, Eisenbach M, Gross H M. Few-shot object detection: a comprehensive survey. IEEE Transactions on Neural Networks and Learning Systems, 2024, 35(9): 11958-11978. doi: 10.1109/TNNLS.2023.3265051
[23]	Wu J X, Liu S T, Huang D, et al. Multi-scale positive sample refinement for few-shot object detection[C]//Computer Vision-ECCV 2020. Cham: Springer, 2020: 456-472.

物种	图片(张)	物种	图片(张)
蚜虫 Aphid	717	褐飞虱 Nilaparvata lugens	556
棉铃虫 Helicoverpa armigera	533	茶翅蝽 Halyomorpha halys	271
草蛉 Lacewing	369	黑食蚜盲蝽 Deraeocoris punctulatus	72
多异瓢虫 Hippodamia variegata	704	牧草盲蝽 Lygus pratensis	107
绿盲蝽 Apolygus lucorum	365	苜蓿盲蝽 Adelphocoris lineolatus	93
麦圆蜘蛛 Penthaleus major	425	白背飞虱 Sogatella furcifera	190
烟盲蝽 Nesidiocoris tenuis	175	白背飞虱 Sogatella furcifera	190

Shot	Baseline	AP	AP₅₀	AP₇₅	AP_S	AP_M	AP_L
10	Faster R-CNN	0	0	0	-	0	0
	YOLOv3	8.1	22.7	4.7	-	16.9	8.6
	DINO	18.9	34.0	15.7	-	1.4	21.5
	Meta R-CNN	15.1	36.6	9.3	-	3.3	15.7
	FSDetView	9.3	24.4	4.5	-	1.3	10.2
	MPSR	27.6	53.3	24.2	-	27.0	27.9
	FSD-MOGA	37.0	57.8	29.9	-	26.4	39.9
30	Faster R-CNN	0.9	5.6	0.1	-	0	1.0
	YOLOv3	6.7	20.8	3.3	-	8.2	7.1
	DINO	8.0	18.0	6.4	-	32.8	8.2
	Meta R-CNN	19.7	47.8	13.2	-	45.0	34.0
	FSDetView	29.2	58.0	27.1	-	4.0	31.3
	MPSR	41.4	70.2	41.0	-	61.5	41.5
	FSD-MOGA	44.0	73.6	45.5	-	75.2	44.0

Shot	MPSR		AP	AP ₅₀	AP ₇₅	AP _S	AP _M	AP _L
Shot	Center Loss	MOGA	AP	AP ₅₀	AP ₇₅	AP _S	AP _M	AP _L
10			27.6	53.3	24.2	-	27.0	27.9
	+		28.9	53.9	26.5	-	23.8	29.8
		+	29.2	54.7	26.9	-	40.6	29.5
	+	+	37.0	57.8	29.9	-	26.4	39.9
30			41.4	70.2	41.0	-	61.5	41.5
	+		42.7	73.1	41.1	-	72.8	42.7
		+	42.9	71.9	43.6	-	72.8	42.9
	+	+	44.0	73.6	45.5	-	75.2	44.0