一种基于YOLO v5的草莓多阶段目标检测方法

A Target Detection Method based on YOLOv5 in Multi-stage of Strawberry Growing Period

草莓是一种流行性较广的高价值水果,在我国种植面积较广。草莓采摘期较长,同一时间可能存在多种形态的果实,针对这一特殊生长习性,提出一种基于YOLO v5的目标检测算法,在准确识别成熟果实的同时,完成多阶段草莓的检测,用于机器人采摘、成熟期预估和生产管理。使用草莓生产环境下图像建立数据集,利用深度学习网络提取草莓生长期各阶段特征。将YOLO v5n、YOLO v5s、YOLO v5m 3种YOLO v5系列的模型和CIoU、EIoU、SIoU、AlphaIoU 4种损失函数计算方法组合,形成了12种算法,在数据集上进行对比试验,结果表明SIoU更适合本研究。将优化后的模型在Jetson Xavier NX和Jetson Nano 2款嵌入式开发板上进行推理效率验证,明确了不同开发板使用的最优模型,Jetson Xavier NX更适合YOLO v5m+SIoU的模型、Jetson Nano更适合YOLO v5s+SIoU的模型,为草莓智能化生产奠定基础。

Strawberries,an epidemic fruit of higher value,are far more widely grown in China.Picking plays an essential role in strawberry production.Robots are an effective way to tackle this problem.In this paper,we proposed a YOLO v5 based object detection algorithm to extract features for each stage of strawberry growing period.By combining three YOLO v5 models,YOLO v5n,YOLO v5s,and YOLO v5m,and four loss function computation methods,CIoU,EIoU,SIoU,and AlphaIoU,12 algorithms were formed and compared experiments were performed on our dataset.The results showed that SIoU was more suitable for this study.The inference efficiency of the optimized model was verified on two embedded development boards,Jetson Xavier NX and Jetson Nano,to meet different smart requirements.The YOLO v5m+SIoU model is more suitable for the Jetson Xavier NX,and the YOLO v5s+SIoU model is more suitable for the Jetson Nano,laying the foundation for intelligent strawberry production.