摘要
为改善现有火焰检测算法参数量大、训练时间长等缺点,本研究提出基于YOLOv4改进的轻量级火焰检测算法。算法以YOLOv4为基本框架,采用MobileNet v3作为主干网络,利用深度可分离卷积替代YOLOv4中颈部网络和检测网络的3×3普通卷积,并将激活函数更换为H-swish函数,构建出一种轻量级火焰检测算法。不仅参数大幅度减少,而且能提升火焰检测精确度,降低火焰漏报率。实验证明,在相同的训练条件下,本研究提出的算法参数量个数降为YOLOv4的18%,训练时间减少44%。当检测相同火焰图像时,与MobileNet v3-DW-YOLOv4算法相比,本研究算法的精确度提升1%,检测速度为每秒46帧,能更好地嵌入到终端设备上进行实时检测。
To improve the existing flame detection algorithms that have such disadvantages as large numbers of model parameters and long training time,this paper proposes an improved lightweight flame detection algorithm based on YOLOv4.Taking YOLOv4 as the basic framework and adopting MobileNet v3 as the backbone network,this algorithm utilizes deep separable convolution to replace the 3×3 ordinary convolution of the neck network and detection network in YOLOv4 and changes its activation function into the H-swish function,thus constructing a lightweight flame detection algorithm model.It not only significantly reduces model parameters,but also improves flame detection accuracy and reduces the leakage rate of flame detection.It is experimentally demonstrated that under the same training conditions,the number of parameters of the proposed model is reduced to 18%of YOLOv4,and the training time is reduced by 44%compared with YOLOv4.Compared with MobileNet v3-DW-YOLOv4 when used to detect the same flame images,the proposed algorithm improves the accuracy by 1%and the detection speed is 46 frames per second.Therefore,it can be embedded in terminal devices for better real-time detection.
作者
王海群
张成君
张怡
WANG Haiqun;ZHANG Chengjun;ZHANG Yi(College of Electrical Engineering,North China University of Science and Technology,Tangshan 063210,China)
出处
《山东科技大学学报（自然科学版）》
CAS
北大核心
2023年第1期91-99,共9页
Journal of Shandong University of Science and Technology(Natural Science)
基金
河北省自然科学基金项目(F2019209553)。
