Smoke Detection Based On The Dual-Channel Convolution Neural Network

There are many fires around the world every year.Once the fire spreads,it is difficult to extinguish them in a short time.Fire control focuses on prevention.If early warning can be issued at the early stage of fire through technical means,many fires can be put out before a large-scale disaster.Smoke is the most obvious feature before the flame burns,so smoke detection can play a role of “preventing the fire from not burning”.Compared with indoor smoke detection,the field smoke detection faces more difficulties and challenges.In the field environment,the spatial range is relatively large,generally several kilometers to more than ten kilometers of monitoring range,and clouds,fog,birds,branches,sunlight,etc.will cause greater interference to the identification of smoke.Smoke is a non-rigid substance,but a translucent one.Its shape and texture characteristics are easy to be interfered by the outside world.The physical characteristics of smoke determine the difficulty of smoke detection.Aiming at the work of smoke detection,this thesis proposes two network models.The first one is a Dual-Channel Convolutional Neural Network(DC-CNN)based on transfer learning.The second is a Dual-channel Convolutional neural Network smoke detection method that fuses Dark Channels of smoke(Dark C-DCN).First of all,for the first network model,in order to solve the problem of insufficient data sets,the transfer learning strategy is used on the first network channel,and the pre-training model is used to generate by training on the large data set Image Net to initialize the weight of the target model;Secondly,the network is set to a dual-channel network structure,and the second network channel is used to directly train the smoke data set,so that it can extract the rich and detailed features of the smoke.Finally,the Concat layer is used to fuse the features to generate a training model to complete the task of smoke detection.For the second network model,a dual-channel residual network method that combines the dark channel features of the smoke image is proposed.The dark channel images are used as another data set of the network model.The network is trained separately to extract the dark channel features of the smoke.The residual block network is introduced into the network structure of the first channel to effectively solve the gradient problem and accelerate the convergence speed of network.The Concat layer is used to fuse features of two channels to increase the comprehensiveness of features,generate training models,and complete the task of smoke detection.In summary,this thesis establishes the smoke detection method based on the dual-channel convolutional neural networks,which can identify smoke accurately in complex environments.The experimental results show that two models have good performance on smoke detection tasks,and the accuracy reaches 99.33% and 99.08% respectively.The methods can successfully detect the smoke in the image and prevent early fires effectively.