YOLO一Sugarcane:用于快速检测复杂背景下 甘蔗植株的轻量级神经网络

打开文本图片集

中图分类号：S566.1 文献标识码：A 文章编号：2095-5553（2025）10-0210-08

Abstract：Inorder tosolvethe problem that theexistinglightweight networkmodel is susceptibletoomissionand misdetection due tosugarcane oclusion，an improved YOLO—Sugarcane lightweight neural network model is proposed. First，the Bneck moduleofMobileNetV3 isused toreplace the MPand ELANmodules withintheYOLOv7—tiny backbone network，whichefectivelyimproves therecallof themodel.Secondlythedepthseparableconvolutionisusedtoplacethe standard 3×3 convolution，which reduces the computational cost of the model and ensures its stable deployment on the sugarcane harvester.Finally，thelightweightchannelattentionmoduleECA（EficientChannelAtention）isintroduced， whichenablesthemodeltoefectivelycapturethefeatureassociationsbetweendiferentchannels，focusonefectivefeatures anddiscardredundant features.Theexperimentalresultsshow that comparedwith theYOLOv7—tinymodel，the floating-point operation amount of the YOLO—Sugarcane model is reduced by 29.18% ，the mAP is increased from 97.10% to 98.18% ，the recall rate is increased from 85.21% to 93.71% ，and the detection speed reaches 91 frames/s.

Keywords： sugarcane；neural networks；attention mechanism；deep learning；lightweight model

0 引言

要原材料。（剩余13588字）