基于YOLOv5s网络的常见菜品识别检测
Common dish identification and detection based on YOLOv5s network
季旭 1宋垲 1冯怡然2
作者信息
- 1. 大连工业大学 机械工程与自动化学院,辽宁 大连 116034
- 2. 大连工业大学 机械工程与自动化学院,辽宁 大连 116034;大连工业大学 辽宁省海洋食品加工技术装备重点实验室,辽宁 大连 116034
- 折叠
摘要
为解决食堂与餐厅的菜品人工结算效率低的问题,基于YOLOv5s网络模型提出几类改进算法,其中,YOLOv5s,YOLOv5s-C2f,YOLOv5s-SE,YOLOv5s-MobileNetV3,YOLOv5x模型的准确率、召回率、均值平均精度分别为 83%,88.6%,89.4%;94.9%,79.1%,87.6%;91.6%,76%,84.9%;88.3%,94.9%,81.5%;93.6%,99.4%,99.4%;每张图片的检测时间分别为0.36,0.29,0.34,0.23,0.98 s.试验结果表明,相较于YOLOv5s模型,YOLOv5s-MobileNetV3模型均值平均精度降低7.9%,检测时间缩短36.12%;YOLOv5s-C2f模型均值平均精度降低1.8%,检测时间缩短19.44%;YOLOv5x模型均值平均精度提升10%,检测时间延长63.27%.YOLOv5s-MobileNetV3模型在维持准确率的同时,大幅缩短检测时间,有效地实现检测效率与性能间的平衡;YOLOv5x模型具有较高准确率,适用于追求高精准度的场合.研究为智能餐饮服务提供技术支持.
Abstract
In order to solve the problem of manual settlement and low efficiency of canteen and restaurant dishes,several types of improved algorithms were proposed based on the YOLOv5s network model,among which the accuracy,recall and average accuracy of YOLOv5s,YOLOv5s-C2f,YOLOv5s-SE,YOLOv5s-MobileNetV3 and YOLOv5x algorithms were 83%,88.6%,89.4%;94.9%,79.1%,87.6%;91.6%,76%,84.9%;88.3%,94.9%,81.5%;93.6%,99.4%,99.4%,respectively;and the detection time of each image was 0.36,0.29,0.34,0.23,and 0.98 s,respectively.Experimental results show that compared with the YOLOv5s algorithm,the average accuracy of the YOLOv5s-MobileNetV3 algorithm is reduced by 7.9%,and the detection time is reduced by 36.12%.The average accuracy of the YOLOv5s-C2f algorithm is reduced by 1.8%,and the detection time is reduced by 19.44%.The average accuracy of the YOLOv5x algorithm is increased by 10%,and the detection time is increased by 63.27%.The YOLOv5s-MobileNetV3 algorithm maintains accuracy while greatly reducing the detection time,effectively achieving a balance between fast detection and performance.The YOLOv5x algorithm has a high accuracy rate and is suitable for applications where high accuracy is required.The research provides technical support for smart food services.
关键词
深度学习/C2f模块/通道注意力机制/YOLOv5s/菜品识别Key words
deep learning/C2f module/channel attention mechanism/YOLOv5s/dish identification引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据