融合多尺度特征和多重注意力的棉田杂草检测研究

打开文本图片集

中图分类号：S562；TP391.4 文献标识码：A 文章编号：2095-5553（2025）10-0138-08

Abstract：Toaddressthe limitations existingcotton field weed detectionmethods，which haveten been criticized for recognizing alimitednumber weed categories exhibiting lowaccuracy，anovelcotton field weed detectionapproach was developed by integrating multi-scale features with multiple atention mechanisms.This method enhanced the feature extractioncapabilitiesthe model byincorporating aneficient multiscaleatention moduleinto the backbonenetwork， without increasing the number parameters or computational cost.Additionally，a progresive feature pyramid was introduced inthe head network turtherenhance featurefusion.Finall，animproved boxregression lossfunction was employedto boost convergence speed positioning accuracy themodel.Experimental results obtained from the CottonWeedDetl2 dataset indicated that the proposed method achieved an average detection accuracy （mAP） 94.6% （20 an F1 score O.754.Compared to the original model，the new method improved mAP by 2.62% ，recall rate by （204 3.16% ， achieved a detection time 65.359ms ，thereby meeting the requirements for real-time monitoring. This approach efectively addressed the challenge accurately detecting weeds incotton fields under natural conditions provides a valuable reference for advancing intelligent weeding technologies in agricultural settings.

KeyWords：cotton field；weed detection；YOLOv8 algorithm；attention mechanism；feature pyramid networks

0 引言

棉花在全球纺织原料中占据核心地位，对农业具有显著影响。（剩余12966字）