基于CNNs和Transformer混合的茶田路径分割模型

打开文本图片集

中图分类号：S571.1；TP391.4 文献标识码：A 文章编号：2095-5553（2026）02-0217-09

Abstract：Theusable partof teacrops is the leaves，which cannot be directlyused foragricultural operations in the air. Currently，thewidelyused dronescannot metthedemand.Therefore，teafieldautomatic navigation technologyapplied to agriculturalunmanned vehicles is becoming increasinglyimportant.To providereal-time navigation path calculation for rapid reasoning in edgecomputing，a new tea field path segmentation model，Tea—CNN—Transformer—YOLO（TCT— YOLO），basedon hybridstrategiesof CNNand Transformer isproposed.Under the backdropof diferentiated agricultural conditions，TCT—YOLO，basedon YOLOv8，implements Transformer-Convolution Blocks（TCB）and （TTB）inastackabledeployment.It exhibitssuperior generalization capabilies intea fieldroadsegmentation tasks，with mIoU improvement of 4.8% .Transformer Blocks in the task may lead to overfiting and gradient explosion，Leaky ReLUactivation functionsand LayerNormalization areadoptedin TCBand TTB，demonstratingoutstanding performance inaccuracy-speedcomparisonswithvariousmodels indownstream tasks.YOLO models’useof Feature PyramidNetworks（FPN）forfeatureextractionoverlookstheproblemofinter-layer informationloss，itisexpectedto improvethem by Gather-and-Distribute（GD）mechanism，fusing，anddistributing features from diferenthierarchical levels，resulting in an total Intersection overUnion improvement of 2.2% ·

Keywors：：tea field path segmentation；instance segmentation；gatherand-distribute（GD）mechanism；edge computing; lightweightmodel

0 引言

中国茶产业的产量与贸易量稳居世界前列，对世界茶产业的发展产生重要的影响[1。（剩余15086字）