改进YOLOv8n的林业害虫检测方法  

Forestry pest detection method based on improved YOLOv8n

作　　者：陈万志 袁航 Chen Wanzhi;Yuan Hang(School of Software,Liaoning Technical University,Huludao 125105,Liaoning,China)

机构地区：[1]辽宁工程技术大学软件学院,辽宁葫芦岛125105

出　　处：《北京林业大学学报》2025年第2期119-131,共13页Journal of Beijing Forestry University

基　　金：辽宁省教育厅科学研究基金面上项目(2021LJKZ0327);辽工程GPU资源支持项目(2024-02)。

摘　　要：【目的】针对现有林业害虫检测方法检测速度慢,检测类别少,小目标害虫检测效果差等问题,提出了一种改进YOLOv8n的林业害虫检测方法。【方法】首先,采用高效多尺度级联注意力特征提取网络EfficientViT作为改进模型的主干网络,降低计算复杂度,提高检测速度;其次,通过构建多尺度自适应特征融合模块DA-C2F提升模型在复杂背景下害虫目标的聚焦能力和识别精度,此外新增的小目标检测头XSH能够进一步提升小目标害虫的检测能力;最后,采用基于最小点距离交并比损失函数MPDIoU作为模型的边界框损失,提升网络收敛速度,进一步增强害虫目标的定位准确率。【结果】改进模型的检测精确率、召回率、平均精度、平均精度均值(mAP50-95)和F_(1)分数分别达到98.6%、95.7%、98.3%、85.6%和0.979,前4者较原模型分别提升了3.9、2.6、2.8、2.5个百分点,F_(1)分数提升了4.4%;检测速度(帧率)达到了95帧/秒,提升了15.9%,优于更轻量级的模型。此外,对比其他检测模型,改进模型对飞蛾类害虫的检测精确率提升了11.2个百分点,并且两种独立飞蛾害虫综合检测的表现也更为优异。【结论】本研究提出的方法对于林业害虫的检测准确度更高,检测速度更快,且对多类别害虫的检测精度更高,改进模型的泛化能力更强。[Objective]In response to the problem of slow speed,narrow categories,and poor detection of small-target pests in existing forestry pest detection methods,a forestry pest detection method based on the improved YOLOv8n was proposed.[Method]Firstly,an efficient multi-scale cascade attention feature extraction network EfficientViT was adopted as the backbone of improved model to reduce computational complexity and enhance detection speed.Secondly,a multi-scale adaptive feature fusion module DA-C2F was constructed to improve the model’s ability to focus on and accurately identify pest targets in complex backgrounds.Additionally,a newly added small-object detection head XSH further enhanced the detection capability for small pest objects.Finally,a minimum point distance IoU loss function MPDIoU was implemented as the bounding box loss for the model,accelerating convergence speed and further improving the accuracy of pest target localization.[Result]The improved model achieved a detection precision of98.6%,recall of 95.7%,mean average precision of 98.3%,mean average precision at different thresholds(mAP50-95) of 85.6%,and an F_1-score of 0.979.These metrics represented improvements of 3.9,2.6,2.82.5 percentage points and 4.4%,respectively over the original model.The detection speed reached 95 frames per second with an improvement of 15.9%,also surpassed that of the lighter-weight model.Furthermore,in comparison with other detection models,the improved model demonstrated an increase of 11.2 percentage points in detection precision for moth pests,and exhibited superior overall performance in detection of two independent moth pest species.[Conclusion]The proposed method has higher detection accuracy and faster detection speed for forestry pests,better detection accuracy for multiple categories of pests,and better generalization ability of the improved model.

关 键 词：深度学习 卷积神经网络(CNN) 林业害虫检测 YOLOv8n 多尺度级联注意力特征提取网络 多尺度自适应特征融合 小目标检测头 

分 类 号：S763.305[农业科学—森林保护学] TP391.41[农业科学—林学]