Research On Visual Recognition Algorithm Of Smoke Density Based On Deep Neural Network

In fires,smoke always appears before flames and is often the main cause of fire deaths.Smoke density is one of the essential characteristics of smoke and an important standard for measuring fire severity and fire detection.Therefore,smoke density measurement is important for fire detection.significance.Existing methods for measuring smoke density based on video images mainly extract features manually,and require external conditions such as atmospheric light and background of the known environment,cannot independently measure smoke density,and are prone to misjudgment.In order to improve the directness and practicability of the smoke density measurement method,in this paper,the unsupervised optical flow model is used to realize the segmentation of the smoke candidate region,and then the smoke density is directly measured in the smoke candidate region through a two-channel deep convolutional neural network.The main research contents and innovations of this method are as follows:(1)Selection of smoke candidate regions.In this paper,a smoke candidate region selection algorithm FPFlow based on deep learning unsupervised optical flow is proposed,which does not require the true value of optical flow with labels,and realizes the segmentation of smoke candidate regions.FPFlow is divided into three parts: a feature pyramid network,an unsupervised training module and an optical flow estimation network.The prior knowledge in traditional optical flow estimation such as brightness constant constraints and smooth assumption constraints is used to construct an unsupervised loss function,and the optical flow consistency before and after calculation is used to estimate occlusion.region,reduce model oscillation by stopping the gradient at the mask of the occluded area,crop the image size to achieve continuous self-supervision to improve network performance,and expand the receptive field of the convolution kernel without reducing the size of the feature map through dilated convolution.Experimental analysis shows that the algorithm can achieve 71.04% mIOU and89.56% PA for the segmentation of smoke candidate regions.Compared with the existing popular unsupervised optical flow methods,the algorithm has a greater advantage in performance.(2)Direct measurement of smoke density.Through the smog optical model,the smog equation between the background image,the current video frame,the environmental parameters and the smog density is obtained,and the corresponding relationship between the smog image and its density value is established according to the smog equation,and then the smoke density data set is established according to the relationship.Construct a two-channel deep convolutional neural network MCNN.In MCNN,1*1 convolution is used for channel data fusion,and skip connections are introduced to make the neural network train deeper,and a self-attention mechanism is also introduced to automatically learn the importance of hidden features.Features are combined to obtain comprehensive measurements.Experimental analysis shows that the MAE of MCNN for measuring smoke density can be as low as 0.0377,which shows better comprehensive performance compared with existing related methods.To sum up,this method realizes the confirmation of smoke candidate regions through the unsupervised optical flow network model,builds the smoke density data set according to the smoke equation,and finally realizes the direct measurement of the smoke density through the dual-channel deep convolutional neural network.detection accuracy and detection efficiency.