基于YOLOv5的草莓轻量化网络检测模型

YOLOv5-based lightweight network model for strawberry detection

随着现代农业技术的发展,草莓生产和采摘的自动化是一个必然的趋势,而草莓目标检测是实现采摘自动化的关键环节。基于YOLOv5目标检测算法,采用ShuffleNet轻量级网络结构替代原模型的特征提取网络,并在骨干网络提取的特征图后加入SE通道方向的注意力机制,结合EIoU和Alpha-IoU损失函数,设计了一个α-EIoU损失函数,给定参数α的值为3,统一指数化IoU损失函数,据此获得更准确的边界框回归和目标检测。改进的模型在草莓小目标数据集上平均检测精度均值达到了97.6%,其中成熟草莓的准确率为99.4%,与YOLOv3、YOLOv4和YOLOv5相比,平均精度均值(mAP)分别提高了5.4%、2.9%和1.1%,该模型识别图像传输帧率为125 fps,比原YOLOv5模型提升了38 fps,该实验模型更适应于移动端部署,为草莓采摘识别的自动化提供了一些理论基础。

With the development of modern agricultural technology,the automation of strawberry production and picking is an inevitable trend,and strawberry object detection is a key link to achieve picking automation.In this paper,based on the YOLOv5 target detection algorithm,the ShuffleNet lightweight network structure is used to replace the feature extraction network of the original model,and an attention mechanism in the direction of the SE channel is added after the feature map extracted by the backbone network.Combining the EIoU and Alpha-IoU loss functions,anα-EIoU loss function was designed,given a value of 3 for the parameterα,to unify the exponentiated IoU loss function,whereby a more accurate bounding box regression and object detection was obtained.The improved model in this paper achieved a mean detection accuracy of 97.6%on the average strawberry small object dataset,with 99.4%accuracy for ripe strawberries,and improved mAP by 5.4%,2.9%and 1.1%compared to YOLOv3,YOLOv4 and YOLOv5 respectively.The model recognises images transmitted at 125 fps,an improvement of 38 fps over the original YOLOv5 model.The experimental model is more adaptable to mobile deployment and provides some theoretical basis for the automation of strawberry picking recognition.