适用于嵌入式设备的轻量级深度学习网络RepYOLO
A lightweight deep learning network RepYOLO for embedded devices
周橄尉 1陈嘉越 1吴佳伟 1赵雅琪 1赵奕凯 1张小英2
作者信息
- 1. 山西农业大学,信息科学与工程学院,山西 太谷 030801
- 2. 山西农业大学,软件学院,山西 太谷 030801
- 折叠
摘要
提出了一种轻量级深度学习网络模型RepYOLO算法,并将其移植到嵌入式设备MCU/MPU中.该网络模型RepYOLO以YOLOv4为基础网络模型,通过改进YOLOv4的主干网络CSPDarkNet为RepBlock结构、在Neck层中引入CBAM注意力机制以及在Head层替换anchor-based检测头为anchor-free检测头并加入ATSS算法减小计算量、提高推理速度和检测精度.结果表明,与原YOLOv4模型相比,网络模型RepYOLO在小麦穗检测上表现出更显著的优势,其精度率、召回率、F1值和平均精度均值分别提高了4.7、3.6、1.5、1.7个百分点;此外,RepYOLO在嵌入式设备MCU/MPU上的推理时间分别降低了37.03%和41.44%.
Abstract
A lightweight deep learning network model RepYOLO algorithm was proposed and transplanted to embedded device MCU/MPU.The network model RepYOLO took YOLOv4 as the base network model.By modifying YOLOv4's backbone network CSPDark-Net to the RepBlock structure,introducing the CBAM attention mechanism in the Neck layer,and replacing the anchor-based detec-tion head with an anchor-free detection head in the head layer along with integrating the ATSS algorithm,the computational load was reduced,and both inference speed and detection accuracy were improved.The experimental results showed that compared with the original YOLOv4 model,the network model RepYOLO showed more significant advantages in wheat spike detection,and its precision rate,recall rate,F1 value and average precision value were increased by 4.7,3.6,1.5 and 1.7 percentage points,respectively.In ad-dition,RepYOLO reduced inference time on embedded devices MCU/MPU by 37.03%and 41.44%,respectively.
关键词
目标检测/深度学习/嵌入式设备/轻量级网络/RepYOLO/小麦穗Key words
object detection/deep learning/embedded device/lightweight network/RepYOLO/wheat spike引用本文复制引用
基金项目
山西省大学生创新创业项目(20220180)
山西省基础研究青年科学研究项目(202203021212414)
出版年
2024
NETL
NSTL
万方数据