Intelligent Fire Detection Technology Based On Convolutional Neural Network

Fire is one of the destructive,fast spreading and uncontrollable natural disasters.The rapid warning of fire can minimize the loss caused by fire.With the development of artificial intelligence,intelligent detection based on vision provides a new method for fire prevention.Compared with the traditional detection method,fire detection based on video has the advantages of wide monitoring range,visual image and rapid alert.In the early stage of fire,smoke often accompanies it.Smoke detection and identification is particularly important in this stage.This thesis combines the static and dynamic characteristics of smoke to solve the problem of smoke detection in reality.The model of convolutional neural network is improved,so as to realize the reliable detection of smoke.The main work of this thesis is as follows:1.This thesis studies the forms of smoke image in different color space and how to extract the moving region of smoke accurately.On the basis of Le Net’s classical convolutional neural network architecture,the features of smoke images in each convolutional layer are visualized.The relationship between feature images extracted by convolution layer is analyzed.2.To solve the problem of difficult classification of smoke and cloud,a multinetwork model fusion architecture is proposed.The feature extractor of VGG16 and Res Net50 network is integrated to better express the spatial information and semantic information of the image.The full connection layer has been reconstructed to distinguish the smoke from the cloud.Compared with the popular smoke detection methods,the multi-network model fusion method has more advantages in accuracy and false alarm rate.3.Because the number of parameters and computation of neural network model is too large,it is difficult to apply in the embedded platform with limited computing resources.This thesis improves the Mobile Net model of lightweight convolutional neural network.Under the condition that the accuracy is basically unchanged,the real-time and reliable smoke detection can be realized.Smoke detection is mainly divided into two stages.In the extraction stage of suspected smoke area,the static and dynamic features of smoke image are combined.In the smoke detection stage,the output layer of Mobile Net model is improved,and sigmoid activation function is adopted to realize the dichotomy of smoke.Experimental results show that the proposed method has high accuracy and practical value.