基于半监督一致性学习的复杂场景枸杞虫害检测OA
Wolfberry pest detection in complex scenarios based on semi-supervised consistency learning
在农业智能化发展背景下,枸杞虫害的精准识别对于保障枸杞产量和提升田间管理效率具有重要意义.然而,受限于虫害目标尺寸小、密集分布、背景复杂以及标注成本高等因素,现有全监督目标检测方法在枸杞虫害检测任务中面临泛化性不足与高度依赖标注样本的问题.为应对这一挑战,提出了针对蓟马、毛跳甲、蚜虫、负泥虫、木虱5类枸杞虫害的半监督检测框架(dual-teacher wolfberry pest semi-supervised detection,DT-WPSSD).DT-WPSSD利用 YOLOv8构建半监督框架并引入双教师分支结构,利用已有的标注样本训练网络并为未标注样本生成伪标签,并将伪标签划分为不可靠、不确定和可靠3种类型,以抑制低质量伪标签传播.此外,构建交叉监督机制以约束不同教师模型预测结果的一致性,并设计伪标签融合策略提升伪标签的质量与稳定性.同时,收集并重构复杂场景下的多尺度枸杞虫害数据集MSPD,对大尺度、中尺度、小尺度和密集场景虫害样本进行系统分析.最后针对模型结构逐渐复杂、计算开销增加的问题开展模型剪枝实验.DT-WPSSD在mAP@0.5指标上分别在MSPD和IP102数据集上达到了88.2%、57.1%,显著优于半监督基线模型.经过模型剪枝,DT-WPSSD参数量下降6 M,浮点运算量下降7.1×109.研究结果表明,DT-WPSSD在复杂场景中提升了模型检测稳定性与小目标识别能力.
In the context of intelligent agriculture,accurate identification of wolfberry pests is important for ensuring wolfberry yield and improving field management efficiency.However,due to small target sizes,dense distributions,complex backgrounds,and high annotation costs,existing fully supervised object detection methods suffer from limited generalization and a strong dependence on labeled samples in wolfberry pest detection.To address this challenge,this study proposes DT-WPSSD(Dual-Teacher Wolfberry Pest Semi-Supervised Detection),a semi-supervised detection framework for five wolfberry pest categories:thrips,flea beetles,aphids,leaf beetles,and psyllids.Built on YOLOv8,DT-WPSSD establishes a semi-supervised framework with a dual-teacher branch architecture.The network is trained on available labeled samples and generates pseudo-labels for unlabeled samples;these pseudo-labels are then categorized into three reliability levels(unreliable,uncertain,and reliable)to suppress error propagation from low-quality pseudo-labels.In addition,a cross-supervision mechanism is constructed to constrain the consistency of predictions from different teacher models,and a pseudo-label fusion strategy is designed to improve the quality and stability of pseudo-labels.Meanwhile,a multi-scale wolfberry pest dataset(MSPD)under complex scenarios was collected and reconstructed,and pest samples at large,medium,and small scales as well as in dense scenes were systematically analyzed.Finally,to address the increasingly complex model structure and rising computational cost,model pruning experiments were conducted.DT-WPSSD achieved mAP@0.5 scores of 88.2%and 57.1%on the MSPD and IP102 datasets,respectively,significantly outperforming semi-supervised baseline models.After pruning,the parameters of DT-WPSSD were reduced by 6 M and floating-point operations by 7.1×109.These results demonstrate that DT-WPSSD improves detection stability and small-target recognition capability in complex scenarios.
陈曦;刘建平;张建华;王健
北方民族大学计算机科学与工程学院,宁夏 银川,750021北方民族大学计算机科学与工程学院,宁夏 银川,750021中国农业科学院农业信息研究所,北京市,100081中国农业科学院农业信息研究所,北京市,100081
农业科技
目标检测半监督学习一致性学习模型剪枝枸杞虫害复杂场景
object detectionsemi-supervised learningconsistency learningmodel pruningwolfberry pestscomplex scenes
《智能化农业装备学报(中英文)》 2026 (3)
109-120,12
宁夏回族自治区全职引进高层次人才(2024BEH04130)国家自然科学基金项目(32460444)宁夏自然科学基金项目(2025AAC050001)宁夏回族自治区青年科技托举人才(22225000502) Ningxia Hui Autonomous Region Full-Time Introduced High-Level Talent Program(2024BEH04130)National Natural Science Foundation of China(32460444)Ningxia Natural Science Foundation Project(2025AAC050001)Ningxia Hui Autonomous Region Youth Science and Technology Talent Support Program(22225000502)
评论