改进YOLOv8的核桃品种动态检测方法

打开文本图片集

中图分类号：TN911.73-34；TP391 文献标识码：A文章编号：1004-373X（2025）15-0110-09

Walnutvarietydynamic detectionmethodbased on improved YOLOv8

LU Wentao， ZHANG Liping， ZHENG Weiqiang， LEI Jiale，PAN Zhichao （School of Mechanical Engineering，XinjiangUniversity，Urumqi 83oo46，China）

Abstract：Anovel lightweight detection algorithm YOLOv8n-EAMis introduced to enhance theaccuracyof walnut variety detectiononconveyorsystems.Byemployingamulti-scaleconvolutionalstructure，andrefiningGhostNet'sstrategyforminimizing redundantfeatures，groupedonvolution，and MobileNet'schannel feature fusion，themodel'score featureextractioncapacityis enhancedsignificantly.Toensurebeterretentionofinformationandscaleconstancyinfeaturemaps，acombinationof average poolingandmaxpolingisutlizedfordownsampling，whichimprovesthemodel'sdetectionaccuracywhiledecreasingfloatingpointoperations.Additionaly，themodelnowincorporatesanMLCA（mixedlocalchanelatention）mechanismwithinitsspatial pyramidpolingframework，enhancingitsadeptnessatseizingsemanticdetailsacrossvariedscalesattheregionofNeckComparative experimentationonaproprietarywalnutdatasetmanifeststhattheYOLOv8n-EAMalgorithmreducesfloating-pointoperations to some extent and shrinks the model size by17.7%，andadvances precisionrate，recallrate，mAP @0.5 ，and mAP@0.5： 0.95 by increments of 3.6%， 1.6% 1.4% ，and 2.8%，respectively，in comparison with the YOLOv8n algorithm. On the Pascal VOC2007dataset，theYOLOv8n-EAMalgorithmachieves relativelyhighaveragedetectionaccuracy incomparisonwith the other algorithms，with mAP@O.5and mA ′@0.5：0.95 of 58.2% and 37.2%，respectively. These results surpass those of the other object detection algorithms，providing technical support for the industrial sorting of walnut varieties.

Keywords：YOLOv8;convolutional function;lightweight network；atention mechanism;object detection;deep learning;automatic sorting;walnut variety detection

0 引言

核桃是一种富含营养的坚果，被称为四大干果之一，有丰富的蛋白质、脂肪等成分，容易被人体吸收。（剩余14255字）