基于EfficientDet-D1的草莓快速检测及分类

Rapid detection and classification of strawberries based on EfficientDet-D1

为了快速识别自然环境下的成熟草莓与未成熟草莓,本研究提出了基于EfficientDet-D1的草莓快速检测及分类方法。该方法具有EfficientNet网络中快速归一化特征加权融合特点,应用该方法与YOLOv3、YOLOv4、Faster-RCNN以及EfficientDet-D0模型进行对比试验,结果显示,YOLOv3、YOLOv4、Faster-RCNN、EfficientDet-D0和EfficientDet-D1等5种算法的平均精度均值(PmA)分别为89.51%、69.02%、96.54%、96.71%、97.50%。试验结果表明,EfficientDet-D1在成熟草莓与未成熟草莓的检测性能均优于其他4种目标检测算法,有较好的泛化性和鲁棒性,且使用模型参数量较小的EfficientNet网络,更适合作用于移动端识别,可实现草莓快速识别中的速度与精度要求。

A target detection algorithm based EfficientDet-D1 model was proposed to meet the requirements of speed and accuracy in the rapid detection and classification of strawberries using the fast normalized feature weighted fusion feature in EfficientNet network to quickly identify the ripe and unripe strawberries in natural environments.The YOLOv3,YOLOv4,Faster-RCNN and EfficientDet-D0 models were used for comparative experiments.The results showed that mean average precision(mAP)of the five algorithms including YOLOv3,YOLOv4,Faster-RCNN,EfficientDet-D0 and EfficientDet-D1 was 89.51%,69.02%,96.54%,96.71%,and 97.50%,respectively.The detection performance of EfficientDet-D1 in the ripe and unripe strawberries is better than that of the other four target detection algorithms,which has better generalization and robustness,By using the EfficientNet network with a small number of model parameters.It is more suitable for mobile identification,and can provide a new solution for the automatic picking technology of the ripe strawberries.