YOLOv7-VSS輕量化橘瓣外觀檢測(cè)模型
打開文本圖片集
DOI：10.16652/j.issn.1004-373x.2025.10.014 引用格式：，等.YOLOv7-VSS輕量化橘瓣外觀檢測(cè)模型[J].現(xiàn)代電子技術(shù)，2025，48（10）：85-91.
中圖分類號(hào)：TN911.23-34文獻(xiàn)標(biāo)識(shí)碼：A
文章編號(hào)：1004-373X（2025）10-0085-07
Abstract：Inallusion to the problems of slow speedandlowaccuracyof orange petal appearance detection in canned citrusproduction，aswellasthehigherparametercountofmainstreamdetectionmodels，alightweightorangesegment appearancedetectionmodel，OO-，ispropsed.Inthemodel，animprovedEcientVietworkisitroducedbysing theHardSwish activation function as the backbone.The mapping similaritybetween diferentdetection heads isreduced by inputingfeaturesatdiferentlevels，whichaleviatesredundantcalculations，andenhancesthenetwork'sfeatureextraction capability bymeansofcascaded groupatention mechanism.Aslim-neck modulethat fuses thepropertiesof standard convolutionanddeepseparableconvolutionisreferenced toreducethesizeofthemodelwhilemaintaining highaccuracy.In orderto furtherreduce the model sizeandspeedupinference spee，SPPCSPC isreplaced with the SPPFstructure.Inorderto alignwiththepositionalcharacteristicsoforangesegmentsinthedataset，theMPDIoUlossfunctionisusedtoimprovethe regresionacuracyof the predicted bounding boxes.The experimentalresults showthatthe proposedorange segmentappearance detection model is63.81%smalerin size compared to YOLOv7，whilerealizing a detectionaccuracyof 96.57% .After deploymentandtestingontheJetsonOrinNano，thebalancebetweenmodelsizeanddetectionaccracyissignificantlymproved compared to similar methods，meeting the requirements of the canned citrus production line.
Keywords：orange segment appearance detection;YOLOv7; lightweight;EficientViT; GSConv; Hard-Swish;MPDIoU
0 引言
柑橘罐頭是我國(guó)柑橘加工產(chǎn)業(yè)中最大宗的商品，其柑橘罐頭加工量占世界總產(chǎn)量的 80% 以上。（剩余8786字）