中圖分類號:S572;S126 文獻標識碼:A文章編號:1007-5119(2025)02-0093-08
1.,青島266101;2.中國農業科學院研究生院,北京100081;3.江西省煙草公司撫州市公司,江西撫州344000;4.江西省煙草公司吉安市公司,)
MobileViT-CBAM Model for Fresh Tobacco Leaf Maturity Recognition Based on Transfer Learning
ZHAO Panzhen12, WANG Songfeng1*, QI Fei3, HU Qiang3, WANG Aihual, LI Yachun4, MENG Lingfeng1, YIN Dong?, DUAN Shijiang4*, WANG Zhisheng4
(1.InstituteofTobacoResearchofChinese AcademyofAgriculturalSciences/KeyLaboratoryofTobaccoBiologyandProcesing
MinistryofAgricultureandRuralAfairs,Qingdao266101,China;2.GraduateSchoolofChineseAcademyofAgriculturalSciences
Beijing10oo81,Chia;3.FuzhoBranchofJngxiProvincialobaccoompanyuzhou34o,Jangxi,China; 4.Jaach ofJiangxiProvincial Tobacco Company,Ji'an343oo9, Jiangxi, China)
Abstract:Toestablishoreeconomicalandeientodestructiveintelligentrecogitionthnologyfortobaccleafmaturitya lightweight network modelMobileVi-CBAMonmobile devices was constructed.Firstly,adatasetwasbuiltbycolecting imagesof the middle andupper leaves of‘Yunyan87withdiferent maturity.The CBAMatentionmechanismmodule wasintroduced into the MobileViTstructuretancethefatureexpresioabilityoffreshtobaccoleafmaturitymages.econdlyteginalactiation function Swish wasreplaced with thesmoother SMUfunction tohelpthemodelconverge faster.Finaly,transferlearning was employedtoimprovethetrainingeficiencyandgeneralizationabilityof temodelandachievetheclasificatiooffreshtobacoleaf maturity in complex field environment.Results showed that MobileViT-CBAM exhibited an accuracy of 9 2 . 8 1 % inmaturity clasificationoffreshtbacoeaves,wichissignificantlysperiortotemodelsofG16,ResNet4,VisionTrasforin Transformer,MobileNetV2,andMobileViT.TheproposedMobileVi-CBAMmodelcaneectivelyidentifythematurityegreeof tobacco leaves,providing technical support for the visual system of inteligent tobacco harvesting equipment.
(eywords: maturity of tobacco fresh leaf; lightweight; classfication model; attention mechanism
適宜的鮮煙葉成熟度是保證煙葉烘烤質量的必要條件,直接影響烤后煙葉的外觀、評吸及香氣品質[1-2]。田間煙葉的成熟度一般分為欠熟、尚熟、適熟、過熟,目前其判定方法以眼看、手摸等感官感受定性識別為主,依據特征包括葉面顏色變化、茸毛脫落程度、葉面發皺程度和成熟斑等[3-4]。此類方法主觀性較強,判別標準不統一,易受操作人員經驗影響。因此,當前研究重點是在煙葉實際生產中提出一種科學、便捷的成熟程度判斷方法,以實現大田鮮煙葉成熟度的客觀、準確、快速識別[5]。
現階段鮮煙葉成熟度識別方法主要分為3類:一是基于化學成分的檢測[6-7];二是基于光譜的檢測[8-10];三是基于機器視覺的檢測。基于化學成分和高光譜的檢測雖然均可以取得較高的準確率和較好的分類結果,但化學成分檢測需要破壞植物組織結構,實時性較差;高光譜檢測需要專業的高光譜成像設備,操作難度較高;而使用基于機器視覺的鮮煙葉成熟度模型識別方法操作簡便,可以實現快速無損檢測,適合應用于煙葉成熟度識別任務[11]。史龍飛等[12]、王杰等[13]和謝濱瑤等[14]分別利用反向傳播神經網絡、極限學習機和支持向量機構建鮮煙葉成熟度識別模型,準確率分別達到 9 3 . 6 7 % /9 6 . 4 3 % 和 9 7 . 5 3 % ,證明傳統機器學習模型在鮮煙葉成熟度識別中的可行性;……