深度学习算法YOLOv4支持下的脐橙树株数识别

Identification of navel orange trees based on deep learning algorithm YOLOv4

针对传统果树单株提取方法流程繁琐、人力成本较高、耗时长等问题,该文提出在复杂自然背景下使用深度学习算法YOLOv4实现脐橙树无人机影像端到端的识别方法,利用网络改进、多尺度融合以及损失函数改进等方法提高复杂背景下小目标的识别率。以赣南某地脐橙果树无人机影像作为数据源,应用YOLOv4深度神经网络分别训练基于不同数据集的模型,通过调整模型阈值参数得到最佳模型,并在强背景、弱背景、稀疏和稠密植株测试集上进行测试。结果表明,基于data748数据集训练且模型阈值为0.6时的YOLO-748模型精确率达91.55%,召回率为98.55%,mAP值为93.38%,F1值达0.985,该模型在复杂自然背景下鲁棒性较好。该方法能为现代农业果园管理提供新的可行方案。

Considering the problems of traditional fruit tree extraction methods such as time-consuming process,cost and difficult implementation,this paper proposed a novel framework based on YOLOv4 algorithm to identify orange trees in the UAV images under complex natural backgrounds.Methods such as scale fusion and loss function improvement improve the recognition rate of small targets in complex backgrounds.Taking the UAV image of navel orange fruit trees in a place in southern Ganzhou as an example,the deep neural network YOLOv4 is used to train models based on different data sets respectively,and the optimal model is obtained by adjusting the model threshold parameters,and tested on strong background,weak background,sparse and dense plant test sets.The results showed that the precision rate,recall rate,mAP value and the F1 value of the YOLO-748 model based on the data748 dataset training with a model threshold of 0.6 was 91.55%,98.55%,93.38%,and 0.985 respectively.The model has better robustness in complex natural backgrounds.This method can provide a new feasible solution for modern agricultural orchard management.