Research On Fire Detection And Recognition Based On Machine Vision

With the continuous extension in the application field of machine vision technology,it has also a certain impact on the intelligent development of disaster detection technology.Among all kinds of disasters,fire belongs to one of the disasters with a wide range of impacts.Since it belongs to the preventable and controllable category,early fire detection and identification are taken as the main research direction.As the main technical support of traditional fire detection,various sensors are seriously affected by the external environment,which promotes the development of machine vision technology in the field of video fire detection and prevention.The occurrence of fire is often accompanied by two forms of things: smoke and flame,so the detection of fire is mainly aimed at these two parts.According to the different characteristics of smoke and flame,corresponding detection and identification algorithms are designed,respectively.The specific research contents and innovations are as follows:1)Aiming at the difficulty of feature extraction caused by hazy border and unstable shape of smoke,a convolutional neural network smoke detection and recognition algorithm based on feature analysis,abbreviated as FAN＿CNN(Feature Analysis Net＿CNN)algorithm,is proposed.The obvious shallow features of smoke are combined with the strong self-learning ability of convolutional neural network for image depth features.FAN＿CNN algorithm consists of two structural layers: feature analysis layer and object recognition layer.The feature analysis layer mainly uses the motion direction feature and color feature of smoke to filter out the interference existing in non-smoke images and smoke images,reducing the calculation order of magnitude and improving the reference of input samples.The object recognition layer uses convolution neural network to extract the depth features of smoke image blocks.Dropout and data enhancement are used to prevent over-fitting problems.Finally,the smoke area in the image is marked according to the coordinates of the image block.Compared with the classical convolution neural network algorithm,FAN＿CNN algorithm has a certain improvement in detection rate.2)Aiming at the problem of insufficient samples caused by no official data set in flame detection research,a flame detection and recognition method based on entropy weighted support vector machine is designed.Based on the motion characteristics of flame,an improved TH-ViBe algorithm is proposed to obtain the suspected flame areas.Then extract the flame texture,gray,area change rate,roundness features to form the feature vector group.Finally,the feature vector group that has completed the importance evaluation is input into the support vector machine to establish an entropy weighted flame detection and recognition model.Compared with traditional support vector machines,the entropy weighted support vector machine flame detection model has more advantages in accuracy and sensitivity.