基于卷积神经网络的烟火智能识别算法研究

相较于传统烟火、烟雾传感器检测方法，基于卷积神经网络算法的烟火检测具有更高的检测精度和效率，并能提供火灾现场全局/局部详细信息。本文提出基于改进YOLOv3算法的烟火识别，应用高斯参数设计损失函数从而建立YOLOv3边界框模型，可预测边界框定位不确定性，减少负样本；为充分利用图像局部特征信息对网络结构进行改进，以实际烟火现场图片为研究对象，完成烟火识别过程计算。利用不同拍摄角度、光照条件自制火焰和烟雾数据集进行测试，结果表明，与传统YOLOv3对比，本文提出的改进YOLOv3算法平均精度提高了4.2%。研究方法将有助于提升智能烟火预警、人员救助和险情跟踪作业水平，最终提升事故灾害的应急能力。

A Convolutional Neural Network Based Algorithm for Intelligent Fire Recognition

Compared with conventional fire and smoke recognition methods, the convolutional neural network (CNN) based algorithms are able to provide many merits including high accuracy and efficiency. Particularly, CNN-based approach can offer detailed information of a fire scene which will not be influenced by environmental interference. In this paper, a fire recognition model based on an improved You Only Look Once Version 3 (YOLOv3) algorithm is proposed. Using Gaussian parameters, we designed a Loss function for establishing bounding box model of the YOLOv3 network architecture which can also predict positioning uncertainty to reduce false positive. Then, in order to make full utilization of the local feature information of targeted images and improve the network architecture, we employed actual fire scene picture as the object to complete the fire identification. Thereafter, we compared the improved and original YOLOv3 algorithm using predefined images which were captured with various shooting angles and illumination conditions of fire. Testing results indicate that the average accuracy has been improved by 4.2% using the modified YOLOv3 algorithm. The presented study should be useful in tasks of smart fire alerting, rescue assistance and emergency tracking which may eventually enhance the response capability for accidents and disasters.