Research On Fire Smoke Recognition Algorithm Based On Deep Learning

The occurrence of fire not only poses a great threat to the safety of personnel,but also causes huge property losses.At present,the detection of fire is mainly divided into two research directions,one is the detection of flame,and the other is the detection of smoke.Among them,compared with flame detection,smoke detection is more difficult,so the current flame detection technology is more mature and the detection effect is better.However,fire is usually detected and a fire has occurred,which does not achieve good fire prevention.For the detection of smoke,most of them are based on sensors and traditional image processing methods.The first method requires a large number of sensors to be installed,and the detection distance is limited.The sensor is easily damaged and aged due to the environment,resulting in great missed detection and false detection.The second method has strong dependence on artificial selection features,and smoke has boundary uncertainty,translucency,background blur and characteristics that are susceptible to other factors,which leads to the accuracy of traditional algorithms for smoke recognition.Greatly affected.In view of the above situation,this paper combines the smoke recognition and deep learning algorithms in the field of machine learning,and combines the existing research results to apply two common models,convolutional neural network model and deep residual network model to smoke.In recognition,it effectively overcomes the shortcomings of the traditional scheme and has been effectively improved in terms of algorithm performance and efficiency.The main results of this paper are as follows:(1)Collecting smoke datasets on the Internet,using the Gaussian mixture model method combined with the color feature method to crop the dataset,and using basic data enhancement methods such as changing image color,saturation,brightness and contrast,artificially increasing noise.To expand the data set,a total of 10,000 positive samples and 10,000 negative samples of 32×24 size were produced.(2)Constructing an 8-layer CNN model,the activation function uses the ELU function,and the model is used for 22-segment video smoke detection,color+ motion+morphology method,Gabor wavelet-based detection,CNN with RELU activation function A series of indicators such as accuracy,false positive rate and false negative rate in smoke identification are compared.(3)Construct a 50-layer ResNet model for smoke identification.During the training process,use the accuracy,recall rate,and F1 score to evaluate the performance.Then use the trained model for 22-segment video smoke detection,and color+motion+form.The method,Gabor wavelet based detection,8-layer CNN network for a series of indicators such as accuracy,false positive rate and false negative rate,the experiment proves that the algorithm has superior performance and performs best in the above indicators.