Research On Fire Detection Method Based On Deep Learning

Fire is one of the most common multiple disasters,which seriously threatens the safety of natural environment and human life and property.It has important research significance and application value to detect and warn fire accurately in real time.Fire detection methods based on traditional image processing and machine learning are usually based on prior knowledge to make artificial feature extraction and threshold determination,which is difficult to meet the real-time requirements,and in the face of complex environment and diverse types of smoke and fire,generalization ability is poor.Combined with the application of deep learning technology in the field of target detection,this study proposes a fire detection algorithm based on deep learning,and implements a fire detection and early warning system based on edge computing,which can carry out real-time monitoring of smoke and fire targets and send out fire early warning in time,so as to achieve effective fire prevention and control.The main research work is summarized as follows:(1)An improved YOLOv5 fire detection algorithm based on wavelet transform is proposed to solve the problem that the interference of pyrotechnic objects in complex environment often leads to misjudgment in fire detection.The algorithm integrates the spectral features extracted from the image by wavelet transform into the convolutional neural network to improve the texture recognition ability of the network.Then,CAC3 module embedded with CA attention mechanism is used to enhance the location information of the network layer after the fusion of wavelet feature,so as to improve the information extraction and positioning ability of the network.Finally,for the specific measure of the real difference between the width and height of bounding box and the balance of samples of smoke and fire with varying degrees of difficulty,the improved ?-EIOU is used as Localization＿Loss to improve the frame orientation accuracy.On the basis of open fire data and self-made fire data,the fire data set is constructed,and the model training and reasoning are carried out.The experimental results show that the improved algorithm achieves a 2.3% improvement in m AP compared with the original YOLOv5 s while the detection speed is guaranteed,achieving a better detection effect on the smoke and fire targets in the fire scene.(2)Based on Huawei IVS1800 edge terminal device,the fire detection and warning system based on edge computing is realized.The system collects the video within the monitoring range in real time through the camera,and uses the method combining GMM foreground extraction and "detection + classification" cascaded network for fire detection and identification.This method uses the improved YOLOv5 algorithm model to obtain the smoke and fire target detection frame,and uses the GMM background modeling method to extract the foreground of the moving target in the video.The overlap threshold between the smoke and fire targets detection frame and the moving target area was determined,and the detection frame without motion change was filtered out.Finally,the detection frame region in the video frame is classified through the Mobile Net V2 network,and the final detection result is output.Once smoke or fire is identified,the alarm information will be generated and uploaded to the user to achieve the purpose of fire warning and timely prevention.Figure [35] table [6] reference [68]...