Research On Lightweight Network Fire And Smoke Detection Method Based On Deep Learning

With the development of the national economy and the rapid development of mechanized factories,the intensive use of electricity and fire in factories and residential buildings has become more and more common.A variety of flammable substances and unsafe use of fire and electricity are more likely to cause fires.Most of the sensor-based fire smoke detection methods need to be in contact with smoke particles or feel the temperature change of the flame to sense the fire.Therefore,it is imperative to study the technical methods of video fire smoke detection and prevention based on deep learning.This paper uses the YOLOv4 and YOLOX target detection algorithms with high singlestage detection efficiency as the Baseline improvement,and proposes FM-YOLO and CSG-YOLO fire smoke detection algorithms.This paper mainly includes the following aspects:Aiming at the deficiencies in the YOLOv4 target detection network,an FM-YOLO target detection algorithm is proposed.FM-YOLO is the starting point.By modifying the convolution module of the CSP(Cross Stage Partial)structure in Backbone,the combination of the inverted residual structure and the depthwise separable convolution is added to the CSP structure to expand the width of the convolution operation and reduce the model volume.,and add the SE(Squeeze-and-Excitation)attention mechanism module and the Drop Path method to form the MBConv(Mobile Inverted Residual Bottleneck Block)module.In the shallow layer of the network,Fused-MBConv(FusedMobile Inverted Residual Bottleneck Block)is used to speed up the model inference speed,and finally,while slightly reducing the amount of model parameters,the target detection accuracy of the model’s fire smoke is further improved.On the Pascal VOC2007 dataset,the parameters of the FM-YOLO model are reduced by 1.4×106compared to the YOLOv4 model,and the detection accuracy m AP is increased by 1.42 percentage points.Aiming at the shortcomings of YOLOX-S feature extraction network convolution operation,the width is low,and the accuracy is lower than that of large target detection networks,the CSG-YOLO algorithm is proposed.The CSG-YOLO algorithm uses a combination of a large-size 7×7 convolution kernel and a depthwise separable convolution in the feature extraction network,and uses Layer Norm,GELU(Gaussian Error Linerar Units)and SE attention mechanism modules to form CSE Block.The feature extraction network after changing the convolution module has a large amount of parameters,and the inference speed is slow.In order to speed up the model inference and reduce the size of the model,this paper adds Ghost Module to the CSP structure.On the fire smoke data set made by ourselves,CSG-YOLO improves the accuracy by 1.63 percentage points than YOLOX-S.There are only a few fire scene data sets available for training in the public data.In this paper,the crawler tool is used to obtain the network fire scene and smoke images,and the Labelimg software is used to complete the production of the Pascal VOC2017 format fire smoke data set after the video is framed.In order to enhance the detection ability of small targets,the Mosaic data enhancement method is used during model training,which effectively improves the problem of insufficient small targets.