Research On Fire Detection Algorithm Based On Deep Learning

Fire is a kind of frequent and dangerous disaster in people's daily life.With the tendency that urban buildings are becoming higher and denser,once fire occurs,people's life and property safety will be under severe threat.It is of great significance to detect fire accurately and timely.Traditional fire detection technologies are mostly based on smoke,light,heat,and other fire parameter sensors to detect fire.However,it is limited by the detection distance,installation position and other factors,while the alarm speed is slow and the detection accuracy is low.Since the beginning of the 21 st century,human society has stepped into the height of information age,and the image fire detection technology has entered people's sight.As it can be combined with the existing video monitoring platform,and it has the advantages of fast response,wide scope and rich alarm information,image fire detection technology has quickly become a research focus in academia.Most of the existing image fire detection technologies are based on manually selected features,which are combined with machine learning classifiers to make classification decisions.However,manually selected features are commonly dependent on prior knowledge of experts,which is not able to describe the essence of flame accurately and comprehensively.The performance of shallow machine learning classifier models is limited to the samples in limited quantities.In addition,for complex classification problems,the expression ability and generalization ability of the model are restricted.This paper has studied the basic principles and related technologies of existing image fire detection algorithm,and some of them were reproduced,in order to point out the shortcomings of existing technologies based on actual detection effects.On the basis of the research on the object detection in the field of deep learning,a new fire detection algorithm based on improved YOLO v3 is proposed,which can automatically extract features,classify flames and locate images in the same deep convolutional neural network.Firstly,the feature extraction network structure is improved to enhance feature reuse and make full use of contextual information,so as to improve the classification accuracy of small-scale targets.Furthermore,the selection method of anchor box is improved to solve the problem of inaccurate location and obtain a higher average Intersection over Union.Finally,by simplifying the loss function,the algorithm is more suitable for fire detection task,and the convergence speed of the neural network is further accelerated.Considering the lack of available open fire data sets,a fire data set containing nearly 14000 fire pictures is built for model training and testing,and all of those pictures are selected from fire videos of multiple scenes and multiple combustion scales.In this paper,the algorithm performance is tested on both image and video data.The results show that the detection speed of the proposed is fast and the precision is high.The proposed has a good detection effect of multiscale fires,at a speed of 26.0frames per second and a precision of 97%.Besides,the false alarm rate can be well suppressed under a variety of complex lightning environments.