摘要
针对现有菜品识别算法存在推理速度慢、参数量大,以及难以部署到资源受限的边缘计算设备上等问题,提出一种基于YOLOv11n的轻量化MD-YOLOv11模型。通过利用MobileNetV4融合主干网络,加速主干网络特征提取;设计精简金字塔网络,构建轻量化C3k2-SNv2模块,提高颈部特征融合的效率;简化检测头结构,提升推理速度;引入MPDIoU损失函数,提升模型边界框收敛速度及检测精度。试验结果表明,MD-YOLOv11模型在菜品数据集上的均值平均精度mAP@0.5=98.55%,当同样部署于树莓派边缘计算设备时,相较于基准模型,MD-YOLOv11在维持精度的同时,检测时间缩短64.58%,参数量下降54.83%,计算量下降73.85%,满足菜品实时检测的现实要求。研究对面向边缘计算的菜品实时识别方法和系统建设具有参考意义。
To address the issues of slow inference speed,large parameter count,and difficulty in deploying to resource-constrained edge computing devices in existing dish recognition algorithms,this paper proposes a lightweight MDYOLOv11 model based on YOLOv11n.By utilizing the MobileNetV4 fused backbone network,the feature extraction of the backbone is accelerated.A streamlined pyramid network is designed,constructing a lightweight C3k2-SNv2 module to improve the efficiency of neck feature fusion.The detection head structure is simplified to enhance inference speed.The MPDIoU loss function is introduced to improve the model's bounding box convergence speed and detection accuracy.Experimental results show that the MD-YOLOv11 model achieves a mean average precision(mAP@0.5)of 98.55%on the dish dataset.When deployed on the same raspberry pi edge computing device,compared to the baseline model,MD-YOLOv11 reduces the detection time by 64.58%,the parameter count by 54.83%,and the computational complexity by 73.85%while maintaining accuracy,meeting the practical requirements for real-time dish detection.This research provides a reference for real-time dish recognition methods and system development oriented towards edge computing.
作者
孙畅
刘海隆
宋梦微
杨锦民
SUN Chang;LIU Hailong;SONG Mengwei;YANG Jinmin(School of Resources and Environment,University of Electronic Science and Technology of China,Chengdu 611731,China)
出处
《包装与食品机械》
北大核心
2025年第6期10-18,共9页
Packaging and Food Machinery
基金
国家重点研发计划项目(2025YFE0102800)
新疆维吾尔自治区科技计划项目(2025E01030)。
