The Research On Fire Detection Algorithm Based On Deep Learning

In recent years,the probability of fire is on the rise,which not only threatens the safety of public life,but also causes serious property losses.Therefore,the research on fire detection is of great significance and value.Traditional fire detectors detect smoke,gas and temperature changes according to sensors,but there are some problems such as limited detection range and single application scenarios.Fire detection techniques based on traditional image processing and machine vision rely on manual selection of static and dynamic characteristics of fire,unable to comprehensively learn all the characteristics of the fire.The deep learning image algorithm is gradually mature,and the deep convolutional neural network can automatically extract flame features in the image to achieve end-to-end detection.At the same time,the detection rate and accuracy are greatly improved.Therefore,this paper studies fire detection based on the deep learning algorithm.The specific research content mainly includes:（1）A fire detection model based on DAI-YOLO is proposed.First,the 512×512 dimensions are used as input,and the traditional convolution in Darknet-53 is replaced with a depthwise separable convolution structure.Secondly,multi-scale detection feature detection is used to increase the shallow detection scale and add a 4-fold up-sampling feature fusion structure to improve the accuracy of flame detection.The intersection ratio is used as a metric function to calculate the distance between the prior frame and the real frame.Finally,ISoftNMS algorithm is adopted before the output of fire image to improve the detection ability of small fire area.The experimental results show that the accuracy of the DAI-YOLO model is 91.2%.Compared with the Faster-RCNN,SSD,and YOLOv3 target detection network models,the accuracy rate is increased by 2.7,9.1,and 5 percentage points,respectively.The average detection accuracy（m AP）is 84.6%,and the detection speed is 35 frames per second.The model not only improves the fire accuracy rate,but also ensures the detection rate,and has certain practicability for fire detection.（2）An anchor-free BM-Center Net fire detection model is proposed.The model first takes Center Net as the basic structure,and reduces the model parameters by introducing the lightweight Mobile Net network as the backbone feature extraction network.Then,Feature Pyramid Networks（FPN）is introduced to improve the feature extraction ability of the model and the detection accuracy of small flame targets.The experimental results show that the average detection accuracy of the model in this chapter for fire targets reaches 75.43%,and the number of frames per second（FPS）is 56 frames per second.The model improves the detection speed and ensures the detection accuracy,and has certain practicability for fire detection.Finally,experiments have proved that the fire detection model based on deep learning proposed in this paper can effectively identify and detect fire areas,which has important value and guiding significance for future research and applications.