基于改进YOLOv5s轻量化模型的红外场景目标检测方法研究
Research on infrared scene target detection method based on improved YOLOv5s lightweight model
刘芷汐 1周春桂 1崔俊杰 1段捷 1岳凯杰1
作者信息
- 1. 中北大学 机电工程学院,太原 030051
- 折叠
摘要
红外技术在防备夜间作战和隐蔽作战中发挥的作用是至关重要的,针对如何平衡红外图像检测精度与轻量化的问题,提出一种基于红外场景下的轻量化目标检测模型M-YOLOv5.该网络模型采用改进的ShuffleBlock模块替换原有的CSP骨干网络.此外,应用轻量级上采样算子CARAFE替换原有上采样模块,在C3 模块中加入SE注意力机制,降低冗余信息,提高特征的区分性和表征能力,重新设计损失函数,E-IoU作为新的损失函数,加快模型收敛速度.在公开数据集FLIR上进行了实验,实验结果表明:改进之后网络模型的平均检测精度达到73.0%,仅降低2.9个百分点,而M-YOLOv5 模型的网络参数数量、理论计算量分别减少40%、39%,模型的推理速度提高 52%,满足部署于边缘设备的需求.
Abstract
Infrared technology plays a crucial role in nighttime and covert operations.To address the issue of balancing the detection accuracy and lightweight design of infrared image detection,a lightweight target detection model called M-Yolov5S is proposed for infrared scenes.This network model replaces the original CSP backbone network with an improved ShuffleBlock module.Additionally,it utilizes the lightweight up-sampling operator CARAFE to replace the original up-sampling module and incorporates SE attention mechanism into the C3 module to reduce redundant information,enhance feature distinctiveness,and representation capability.The loss function is redesigned,with E-IoU as the new loss function,which accelerates model convergence.Experimental tests conducted on the FLIR public dataset show that the improved network model achieves an average detection accuracy of 73.0%,with only a 2.9 percentage point decrease compared to the baseline Yolov5 model.Furthermore,M-YOLOv5S reduces the number of network parameters and theoretical computation by 40% and 39%,respectively,while improving the model's inference speed by 52%,making it suitable for deployment on edge devices.
关键词
红外目标检测/轻量化模型/YOLOv5s/CARAFE/注意力机制/损失函数Key words
infrared target detection/lightweight model/YOLOv5s/CARAFE/attention mechanism/loss function引用本文复制引用
基金项目
中北大学研究生科技立项项目(20221804)
山西省重点实验室开放基金(GDZBKKX-15)
出版年
2024
NETL
NSTL
万方数据