基于改进YOLOv5的橘瓣检测方法

Orange flap detection method based on improved YOLOv5

针对柑橘罐头厂生产橘子罐头时人工分类橘瓣效率差的问题,本文提出一种基于深度学习的橘瓣检测方法。以柑橘罐头厂生产线上的橘瓣为对象,实地采集图像并制作了一个具有2 500张图像的橘瓣数据集;在YOLOv5s模型的基础上针对橘瓣检测对象多、遮挡大的特点,融合注意力机制并改进了损失函数,得到改进的YOLOv5s模型。试验结果表明,在该模型上橘瓣检测的平均精度达到93.7%,单张图像检测耗时25 ms,基于改进YOLOv5s模型的橘瓣检测方法能够满足工厂生产线的实际应用需求。本方法可以为橘瓣自动化分类设备提供高精度的视觉指导。

Aiming at the problem of poor efficiency of manual classification of orange flaps in orange canning factories,a method of orange flaps detection based on deep learning was proposed.In this paper,a data set with 2500 images was made by taking the images from the production line of citrus canning plant as the object.Based on the YOLOv5s model,the attention mechanism is fused and the loss function is improved to achieve the improved YOLOv5s model,in which the average accuracy of the orange lobe detection task reaches 93.7%and the detection time of a single image is 25 ms.The experimental results show that the orange detection method based on the improved YOLOv5s model can meet the actual application requirements of the factory production line.This method can provide high-precision visual guidance for orange petal automatic sorting equipment.