

中圖分類號:S238 文獻標識碼:A 文章編號:2095-5553(2025)07-0138-07
Abstract:Inordertoaccuratelyidentifythemainveinoftobacoleaf,realizemechanicalgraspingandreducetherateof grasping damage,animprovedlightweightobaccoleaf mainveinsrecognitionmodelbasedonYOLOv7—tinywasproposedFirstly, theoriginaltrunk featureextractionnetwork isreplacedbyamore lightweight MobileNetV3basedonYOLOv7—tinynetwork, thedefaulth-swishactivationfunctioninthemoduleisreplacedbyReLUactivationfunction.Then,thecommonconvolutionof theneckisreplacedbyalightweightGSConvandaSlim—Neckdesignisadoptedtoompressthechannelofthemodelandeliminate theredundantfeatureredundancyinordertolightenthenetworkstructure.Atlast,theSIoUlossfunctionwas introducedtoreduce thelossvalueofthemodelandenhancethefusionabilityof themodeltothemainveinoftobaco.Theresultsshowedthatthemap value of the improved model on the tobacco leaf dataset was 91.3% ,at a cost of only 1.6% loss,the parameter quantity was reduced by 51.1% compared with the original model,and the computational load was 4.3G ,only 32.6% of the original model (13.2G). Compared with YOLOv5—s ( 16.5G ,YOLOv6—n(11.4G),Yolox—s (26.8G),YOLOv8—n(8.7G),and YOLOv9—t (7.7G),allofthemwereimproved.Theimprovedmodelcanbedeployedinthemarginalequipmentwithscarcecomputing resources,which provides some technical support for the mechanized harvesting of tobacco leaves.
Keywords:tobacco leaf main veins;lightweight;mechanized harvesting;accurate identification;marginal deployment
0 引言
有重要影響。隨著勞動力成本增長,煙葉生產正陷入缺少勞動力的困難局面,以替代人力為目的的煙葉生產機械化是現代煙葉農業發展的重要課題[1]。煙葉生煙葉是煙草工業的主要原料,對全球經濟發展具產條件的復雜性,具體表現在煙田的不規整性以及煙葉的易損性,導致機械化收獲時難免對煙葉造成一定的損害。為更好地解決上述問題,在煙葉生產機械上部署目標檢測模型,對煙葉的抓取部位實現精準抓取是降低煙葉破損率的有效措施之一,這要求對煙葉的主脈部位進行準確識別。近年來,農業領域應用圖像識別已成為一種趨勢,在煙葉的智能分級、品質檢測和成熟度檢測環節上已被廣泛使用,大多數以煙葉顏色、形狀、紋理特征與機器學習相結合的方法實現[2-4]。隨著計算機視覺的不斷發展,計算機深度學習憑借高速度和高精確度,對煙葉分組和識別的研究越來越多,但仍存在復雜煙葉主脈分析過程自動化低,在遮擋的情況下煙葉主脈難以被識別等問題[5,6]。
農業圖像識別的關鍵是尋找更強大的表征,只捕獲圖像中對于給定任務最顯著的屬性,從而提高模型的性能。YOLO是目前階段最有代表性的單階段檢測器,相比于R一CNN、FasterR一CNN等雙階段目標檢測算法擁有更快的檢測速度[7,8],更適用于農業工程領域[9]。2015年Redmon等[10]首次提出單階段目標檢測的概念,將目標檢測看作回歸問題,大大提高了目標檢測的效率,此后又依據FPN的思想,利用多尺度的特征圖來提高小物體檢測的精度,并在Backbone特征提取網絡中加人了殘差模塊,使模型提取到更深層次的特征[11-13]。……