摘要
为了在变电站的低计算能力设备上部署火灾检测算法,通过多种方式的结合改进YOLOv3的网络结构,实现准确而快速的火灾检测。鉴于火灾图像数据集不足以训练深度神经网络,通过多种手段收集火灾图像,自建火灾图像数据集,并基于线上数据增强方法,进一步扩充数据集;鉴于原YOLOv3网络参数众多,引入MobileNetv3-Large主干网络替换原DarkNet53主干网络来降低网络复杂度,并通过在预测网络部分引入Inverted-bneck-shortcut结构实现多尺度特征图的融合预测;进一步通过引入锚框聚类优化、随机带泄漏修正线性单元(randomized leaky rectified linear unit,RLReLU)激活函数改进网络,提升算法的检测精度。实验结果表明,所提改进YOLOv3火灾检测模型的大小近似为原YOLOv3模型的1/3,推断速度提高了近12%,并且算法的平均识别精度提高了近10%,说明所提改进YOLOv3变电站火灾检测算法能较为快速和准确地识别并定位图像中的火焰。
In order to deploy the fire detection algorithm on the substation equipment with low computing power,multiple technologies are combined to improve the structure of YOLOv3 network,so as to realize accurate and fast fire detection of the substation.I view of the fact that the existing fire image datasets are not enough to train the deep neural network,a home-made fire image dataset is established by means of collecting various fire images,and is augmented by the online data enhancement method.Aiming at a problem of numerous parameters of the original YOLOv3 network,the MobileNetv3-Large backbone network is introduced to replace the original DarkNet53 backbone network to reduce complexity of the network.Meanwhile,the Inverted-bneck-shortcut structure is introduced into the prediction network to achieve multi-scale fusion prediction.Furthermore,by using the anchor box clustering method and randomized leaky rectified linear unit(RLReLU)activation function,the detection accuracy of the algorithm is improved.The experimental results show that the size of the proposed fire detection model is decreased to about 1/3 of the original YOLOv3 model,the inference speed is increased near 12%,and the average recognition accuracy is increased near 10%,which indicates that the proposed improved YOLOv3 substation fire detection algorithm can identify and locate the flames in the images quickly and accurately.
作者
张永
刘明一
许庶威
马宜龙
钱惠敏
ZHANG Yong;LIU Mingyi;XU Shuwei;MA Yilong;QIAN Huimin(State Grid Zhejiang Electric Power Co.,Ltd.,Hangzhou,Zhejiang 310007,China;College of Energy and Electrical Engineering,Hohai University,Nanjing,Jiangsu 211106,China)
出处
《广东电力》
2021年第11期123-132,共10页
Guangdong Electric Power
基金
国家自然科学基金项目(61573001)。
