摘要
为解决现有火灾检测算法无法同时满足高检测率、低误报率以及高实时性的检测需求的问题,提出了一种基于卷积神经网络的改进YOLO v3目标检测算法,通过深度卷积神经网络自动提取火焰特征对全图进行多尺度特征图预测.首先,针对网络公开火灾数据集数量较少、场景种类受限、火焰尺度单一等问题,自建了一个包含13573张火灾图片的火灾数据集用于对模型进行训练和测试,其中训练集图片10014张,测试集图片3559张.接着,为了提升网络对于多尺度目标(尤其是小尺度目标)火焰的特征提取效果,通过在原YOLO v3的特征提取网络Darknet-53中嵌入空洞卷积模块以充分利用上下文信息,扩增感受野的同时保证不丢失特征图的分辨率.此外,在特征提取网络中加入DenseNet密集型连接网络结构单元,以增强特征复用,同时缓解深度卷积神经网络在特征传播过程中的梯度消失问题.该改进的特征提取网络相比原网络层数进一步加深,网络参数量显著减少.结合火灾检测任务需求实际,简化了损失函数,加快了网络的收敛速度.实验结果表明:该算法检测速度快,检测精度高,不仅能够实时检测大尺度火焰,对于火灾发生初期的小尺度火焰也同样检测灵敏,其检测速度可达26.0帧/s,精确率可达97%,且在多种复杂光照环境下均能良好地抑制误报.
Existing fire detection algorithms have exhibited difficulty in simultaneously meeting the requirements of high detection rate,low false alarm rate,and high real-time ability.To solve this problem,an improved you only look once(YOLO)v3 object detection algorithm based on a deep convolutional neural network is proposed.Fire features were extracted by the network to predict fires with multiscale feature maps.First,to solve the problem of the low number of open fire datasets,the limited types of scenarios and the single size of fires,a dataset including 13573 fire pictures was built,which was further used to train and test the proposed model.The training set included 10014 pictures,while the test set included 3559 pictures.To enhance feature extraction towards multiscale fires(especially small-scale fires)and to take advantage of the contextual information,dilated convolutional modules were embedded in the Darknet-53 feature extraction network of the original YOLO v3.This expanded the receptive field without the loss of feature map resolution.In addition,some intensive DenseNet network units were added to improve feature reuse,thereby helping to resolve the vanishing gradient problem during feature propagation of the deep convolutional neural network.The improved network was deeper and the parameter size was smaller than in the original algorithm.Considering the actual demands of fire detection,the loss function was simplified,which further accelerated the convergence rate of the network.Results showed that the detection speed of the proposed algorithm was fast and precision was high.The proposed algorithm was skillful in multiscale fire detection with a speed of 26.0 frames per second and a precision of 97%.Moreover,the false alarm rate was well-suppressed under a variety of complex lightning environments.
作者
张为
魏晶晶
Zhang Wei;Wei Jingjing(School of Microelectronics,Tianjin University,Tianjin 300072,China)
出处
《天津大学学报（自然科学与工程技术版）》
EI
CSCD
北大核心
2020年第9期976-983,共8页
Journal of Tianjin University：Science and Technology
基金
公安部技术研究计划资助项目(2017JSYJC35).
关键词
目标检测
火灾检测
空洞卷积
实时检测
object detection
fire detection
dilated convolution
real-time detection
作者简介
张为(1975-),男,博士,教授;通信作者:张为,tjuzhangwei@tju.edu.cn.
