Real-time litchi detection in complex orchard environments: A portable, low-energy edge computing approach for enhanced automated harvesting  

作　　者：Zeyu Jiao Kai Huang Qun Wang Zhenyu Zhong Yingjie Cai 

机构地区：[1]Guangdong Key Laboratory of Modern Control Technology,Institute of Intelligent Manufacturing,Guangdong Academy of Sciences,Guangzhou,China [2]School of Economics and Management,Beihang University,Beijing,China [3]Department of Electronic Engineering,The Chinese University of Hong Kong,Hong Kong

出　　处：《Artificial Intelligence in Agriculture》2024年第1期13-22,共10页农业人工智能（英文）

基　　金：supported by the GuangDong Basic and Applied Basic Research Foundation(Grant No.2022A1515110007);the Natural Science Foundation of Guangdong Province,China(Grant No.2023A1515012869);the Guangzhou Science and Technology Plan Project(Grant No.202007040007);the GDAS’Project of Science and Technology Development(Grant Nos.2022GDASZH-2022010108,2021GDASYL-20210103090).

摘　　要：Litchi,a succulent and perishable fruit,presents a narrow annual harvest window of under two weeks.The advent of smart agriculture has driven the adoption of visually-guided,automated litchi harvesting techniques.However,conventional approaches typically rely on laboratory-based,high-performance computing equipment,which presents challenges in terms of size,energy consumption,and practical application within litchi orchards.To address these limitations,we propose a real-time litchi detection methodology for complex environments,utilizing portable,low-energy edge computing devices.Initially,the litchi orchard imagery is collected to enhance data generalization.Subsequently,a convolutional neural network(CNN)-based single-stage detector,YOLOx,is constructed to accurately pinpoint litchi fruit locations within the images.To facilitate deployment on portable,low-energy edge devices,we employed channel pruning and layer pruning algorithms to compress the trained model,reducing its size and parameters.Additionally,the knowledge distillation technique is harnessed to fine-tune the network.Experimental findings demonstrated that our proposed method achieved a 97.1%compression rate,yielding a compact litchi detection model of a mere 6.9 MB,while maintaining 94.9%average precision and 97.2%average recall.Processing 99 frames per second(FPS),the method exhibited a 1.8-fold increase in speed compared to the unprocessed model.Consequently,our approach can be readily integrated into portable,low-computational automatic harvesting equipment,ensuring real-time,precise litchi detection within orchard settings.

关 键 词：Litchi detection Automated harvesting Edge computing Neural networks Model compression 

分 类 号：S667.1[农业科学—果树学]