Research On Fire Flame And Smoke Recognition Algorithm Based On Deep Learning

Fires have the characteristics of high destructiveness and rapid spread.Therefore,new challenges have been raised for real-time fire detection and effective prevention.Although sensors such as temperature and smoke can detect fires,real-time performance cannot be guaranteed.Traditional fire image detection algorithms mostly use flames as the detection target.Normally,flames will only be generated in the middle of the fire,so when flames are detected,The fire has already occurred,making it impossible to prevent and control the first time.In response to the above problems,this paper proposes an improved YOLOv3 fire flame and smoke detection algorithm.The main research contents are as follows:(1)Use data enhancement methods such as rotation,sharpening,and cropping to preprocess the data.Use the above methods to expand the data set to improve the generalization ability of the model,thereby avoiding overfitting of the model.The expanded data set contains two types of targets,flame and smoke,with a total of 5000 images(4000 images for training,500 images for testing,and 500 images for verification),and then manually label 5000 images containing targets according to the VOC data set format Image.(2)Based on the self-made flame and smoke data set,through theoretical analysis,the SSD network model and the YOLOv3 network model in the one-stage algorithm are selected for comparison experiments.Configure the network hardware environment,build the model and train.Based on the experimental verification results,the YOLOv3 network model is better than the SSD network model in terms of accuracy and speed.Therefore,this paper selects the YOLOv3 network model as the fire flame and smoke detection model.(3)Improve the YOLOv3 network model selected in this article.First select MobileNetV1 to replace Darknet-53,the backbone network of YOLOv3,to change the convolution method of the original network,reduce the parameters required for network operations,and the amount of network calculations,and avoid the problem of excessive memory usage during network training.Next,in order to balance the accuracy loss caused by the increase in speed and the reduction in the amount of parameters,the Mish function is selected as the activation function of the network,which effectively enhances the semantic information of flame and smoke in the network,and improves the accuracy of flame and smoke target detection.Secondly,in the three output predictions of MoblienetV1,the lightweight RFB module is used for re-adjustment,and the feature extraction ability and robustness of the network are enhanced by increasing the receptive field.Finally,due to the irregular shape of flames and smoke,and the complicated environment,the model after introducing the CBAM module can pay more attention to the target object itself,and the positioning of flames and smoke is more accurate.The experimental study proves that compared with the original YOLOV3 network model,the number of parameters of the improved target detection algorithm proposed in this paper is reduced by nearly half,and the MAP value is increased by 0.13%.At the same time,the detection speed is also improved.The average detection time per image is shortened by 0.013s,and real-time detection of flame and smoke targets is realized.