改进YOLOv5s的自然场景下茶叶嫩芽检测

打开文本图片集

中图分类号：TP391.4；S57.1 文献标识码：A 文章编号：2095-5553（2025）08-0103-09

Abstract：The plucking oftendertea buds is a crucial stage in tea production，and intellgenttea harvesting reliesonadeep learning-basedteabuddetectionalgorithmasatechnical foundation.Toenhance thespeedandaccuracyof teabud detection，this paper proposes an improved detection algorithm TN—YOLOv5s based on the YOLOv5s network model. First，GhostConv，adepthwise separableconvolution fromthe GhostNetnetwork structure，is introduced toreplace the ordinaryconvolution layers in the feature extraction andfusion networksoftheoriginalYOLOv5s model.Second，theCA spatial atention mechanism isaddedattheendof themodel's feature extraction network.Third，theSIoU_Loss is employed as the regresson lossfunction in place of CIoU_Loss.Finally，Soft—NMS is used to replace NMS.The results of the study show that the improved model achieved 7.1% ， 5.9% ，and 6.4% higher model precision，recall，and average precision values，respectively，and the weight size decreased from 13.7 MB to 7.48MB compared to the original YOLOv5s algorithmonthecustomteabudsdataset.Furthermore，whencomparedtocurrntmainstreamdetectionalgorithms，the improved model shows beter performancein detectionaccuracy，model size，and detectionspeed.The improvedmodel reducestheleakagedetectionrateof obscuredteabuds，andenable saccurateandrapiddetectionof teabuds indiferent scenarios，providing atechnical foundation for the development of tea-picking robot technology and equipment.

Keywords：tea buds；natural scenes；YOLOv5s；object detection;oclusior

0 引言

茶叶嫩芽采摘是茶叶生产中的重要环节之一，高质量的采摘对于提升茶叶产量、质量和经济效益具有重要意义[1。（剩余14593字）