基于多任务学习的桃园环境检测方法研究

打开文本图片集

中图分类号：TP391.4；S662.1 文献标识码：A 文章编号：2095-5553（2025）10-0146-08

Abstract：The peach orchard scene is complex.To asist agricultural robots in beter perceiving theenvironmentof peach orchardsandquicklyandaccuratelyidentifying peachesand pathways withinthem，andaiming atthe practical problems such as singletask，low detection accuracyand slow reasoning sped ofthecurrent model，an eficient multi-task learing network named MTL—YOLO is proposedby improving YOLOv5n.The network simultaneously accomplishes the tasks of objectdetectionandsemanticsegmentation.Firstly，anaditionaldetectionheadfordrivableareasegmentationisadded to YOLOv5n to detect peaches and pathways within the orchards.Secondly，a lightweight ShufeNet V2 is employed as the backbone network of MTL—YOLO，which reduces the computational complexitywhileensuring detection acuracy. Furthermore，the RepNCSPELAN4 module isembedded inthe Neck partof the model，replacing the original C3 module， toenhance feature extraction capabilitiesand further reduce computational complexity.Finally，an adaptive loss weight adjustment method suitable formulti-task models is proposed toavoid thecumbersome processofmanuallyoptimizing loss weights forthetwo tasksand strengthenthecorrelation between themduring training.Experimentalresults showthatthe improved MTL—YOLO achieves an object detection accuracy of 84.7% ，an increase from the original algorithm's 82.1% .Moreover，the semantic segmentation accuracy isincreased by 0.3% and by 2.5% ，compared to mainstream

Mask R—CNN and YOLACT algorithms，respectively. The real-time detection speed of the model reaches 110f/s Keywords：target detection；semantic segmentation；multi-task learning；lightweight；adaptive lossfunction

0 引言

我国是桃子产量最多的国家。（剩余11563字）