融合動(dòng)態(tài)稀疏注意力的茶葉分類檢測(cè)

打開文本圖片集

中圖分類號(hào)：TP391.41；S225.99 文獻(xiàn)標(biāo)志碼：A 文章編號(hào)：1008-0864（2025）07-0111-11

ClassificationandDetectionofTeaBasedonDynamic Sparse Attention

ZHANG Shihao 1，2 ，XIAYuxin2，WUWendou3，XIE Jin 1 ，CHEN Xiao4， SHIHaotian1，F(xiàn)AN Zongyu1，WANG Baijuan1*

（1.Colegeofaince，YunnanAgriculturalUiversityKunming6Oina;.CollgeofechanicalandEltial

Engineering，YunnanAgriculturalUniversity，Kunming 65OO1，China;3.CollgeofBigData，unnanAgriculturalUnivesity， Kunming 650201，China;4.Yunnan Province University Intelligent Organic TeaGardenConstructionPrimaryLaboratory， Yunnan Agricultural University，Kunming 65O2O1，China）

Abstract：Inorder to effectively address thechallenge of precisely detectingand selecting tea leaves witha teapicking robot，an improved YOLOv7 algorithm based on double-layer routing dynamic sparse atention mechanism andFasterNet wasproposed to realize the classificationand detectionoffresh tealeaves.The algorithm implemented involves replacing the original network’s structure with PConvand FasterNet methodologies.This replacement was aimed atdecreasing the numberof floating-point operations and enhancing their eficiency.Additionally，adynamic sparse attention mechanism，which was based on a double-layer routing approach，was incorporated into the neck layer.This adition ensured greater flexibility in computing alocation and content perception.To expedite convergence，enhance regression accuracy，and minimize false detection，the loss function was substituted with effcient intersection over union（EIoU）during the detectionprocess.The resultsshowed that，comparedwith the originalYOLOv7，the model generated bythe improved algorithm improved the accuracy by4.8 percentage point， the recall rateby5.3percentagepoint，thebalancescore by5.Opercentage point，andthe meanaverage precision （mAP）value by2.6percentage point.In the external verification，the numberof floating-point operations was reduced by 15.1 G，the frame per second was increased by 5.52% ，and the mAP value wasincreased by 2.4 percentage point.The improved model could notonly classifyand detectfresh tea leaves eficientlyandaccurately， but also hadthe characteristicsof highrecognitionrate，lowcomputationand fastdetection.Aboveresultslaida foundation for the realization of tea picking robot in Yunnan plateau.

KeyWords：tea；vision transformer with Bi-level routing attention；accurate detection；FasterNet； YOLOv7

茶葉作為云南傳統(tǒng)的特色支柱產(chǎn)業(yè)，在建設(shè)云南高原特色現(xiàn)代農(nóng)業(yè)中扮演著重要角色1。（剩余13303字）