作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程

• •    

PestDA-YOLO:跨域少样本害虫检测方法

  • 发布日期:2026-11-20

PestDA-YOLO: A Method for Cross-Domain Few-Shot Pest Detection

  • Published:2026-11-20

摘要: 在农业害虫监测中,不同黄色粘虫板监测场景之间存在明显的分布差异,源域模型在目标域仅使用少量人工标注时难以充分适应新监测场景,导致目标域检测性能下降。针对上述问题,提出跨域少样本害虫检测方法PestDA-YOLO。该方法在YOLOv11s的骨干特征提取网络(Backbone)和多尺度特征融合网络(Neck)中嵌入域适配模块(Domain Adapter,DA),利用轻量瓶颈残差分支对不同层级特征进行目标域适配;在P3/P4分类路径中设置检测头集成特征增强模块(Detection-Head Integrated Feature Enhancement,DIFE),融合局部细节与全局统计信息,增强小尺度害虫的特征表达,为目标类别判定提供更充分的特征信息;同时构建伪标签部分标注补全策略(Pseudo-label Partial Annotation Completion,PPAC),由冻结教师生成并筛选候选预测,补充目标域K-shot训练图像中未被人工标注覆盖的害虫实例,缓解错误背景监督。模型先利用源域多样本数据进行训练,再使用目标域K-shot标注完成微调,并在番茄粘虫板与棉田粘虫板之间开展双向迁移实验。实验结果表明,PestDA-YOLO在每类30、50和100个目标域标注实例的设置下,交并比(Intersection over Union,IoU)阈值为0.5时的平均精度均值(mAP50)均为所有对比方法中最高。相较于YOLOv11s基线,在番茄目标域的30/50/100-shot设置下,mAP50分别提高10.7/11.1/8.3个百分点;在棉田目标域的相应设置下,分别提高6.6/4.7/9.3个百分点。模型包含11.44×106个参数,计算量为22.36十亿次浮点运算量(GFLOPs),在两个目标域上的前向推理速度均约为54.5帧/s(FPS)。综上,PestDA-YOLO能够在有限目标域人工标注条件下有效缓解跨监测域特征适应和监督信息不足问题,在保持较低计算开销和实时推理能力的同时,提高黄色粘虫板跨监测域少样本害虫检测性能。

Abstract: In agricultural pest monitoring, clear distribution differences exist among yellow sticky trap monitoring scenarios. A source-domain model cannot fully adapt to a new monitoring scenario when only a small number of manual annotations are available in the target domain. This limitation reduces target-domain detection performance. To address this problem, this paper proposes PestDA-YOLO, a cross-domain few-shot pest detection method. The method integrates a Domain Adapter (DA) into the backbone and neck of YOLOv11s. The DA employs lightweight bottleneck residual branches to perform domain-related feature modulation on hierarchical features. The method also introduces a Detection-Head Integrated Feature Enhancement module (DIFE) into the P3/P4 classification paths. DIFE integrates local details and global statistical information to enhance small-scale pest feature representations for category prediction. In addition, this paper develops a Pseudo-label Partial Annotation Completion strategy (PPAC). A frozen teacher generates and filters candidate predictions. The selected predictions supplement pest instances that are not covered by manual annotations in target-domain K-shot training images. This process alleviates erroneous background supervision. The training procedure first uses multi-sample source-domain data to train the model and then uses target-domain K-shot annotations to fine-tune it. This paper conducts bidirectional transfer experiments between tomato and cotton yellow sticky traps. The experimental results show that PestDA-YOLO achieves the highest mean average precision at an intersection-over-union threshold of 0.5 (mAP50) among all compared methods under the settings of 30, 50, and 100 annotated target-domain instances per class. Compared with the YOLOv11s baseline, PestDA-YOLO improves mAP50 by 10.7/11.1/8.3 percentage points under the 30/50/100-shot settings in the tomato target domain, respectively, and by 6.6/4.7/9.3 percentage points under the corresponding settings in the cotton target domain. The model contains 11.44×106 parameters and requires 22.36 billion floating-point operations (GFLOPs). Its forward inference speed is approximately 54.5 frames per second (FPS) in both target domains. In summary, PestDA-YOLO effectively addresses the challenges of cross-domain feature adaptation and insufficient supervision under limited target-domain annotations, while maintaining low computational overhead and real-time inference capability, thereby improving few-shot pest detection performance across yellow sticky trap monitoring domains.