Research On Tunnel Fire Warning Based On Intelligent Image Processing Technology

Once the tunnel fire occurs,the fire is rapid,difficult to rescue,great harm.Traditional tunnel fire detection methods,such as temperature-sensing cables,need to reach a certain temperature before alarm,which takes a long time.Deep learning and other methods are used to quickly detect the emergence of flames in the tunnel in real time,so as to nip the fire in the bud.Deep learning,which automatically mining and analyzing features from a deeper level,is a new idea in flame detection.In order to solve the problems of slow speed and high false detection rate of traditional tunnel fire detection methods,this paper first proposes a YOLOv5s-SRGAN cascade network algorithm for small flame targets.Low-resolution flame images are obtained according to prior information,and the super-resolution reconstruction algorithm based on generative adversance network is used to restore pixels.The generated super-resolution features were fused with the original small flame target,and then feature extraction and discrimination were carried out by YOLOv5 s network.In order to make YOLOv5s pay more attention to the flame area and reduce background interference,the attention mechanism is introduced.CBAM is introduced in the trunk network and SE-Net is introduced in the neck network respectively.Finally,the YOLOv5 s loss function is constructed.The classification loss function of YOLOv5 s is combined with the gradient harmonized mechanism to solve the problem of extreme imbalance between positive and negative samples,accelerate the model convergence and optimize the model performance.To solve the re-identification problem of pedestrian detection,the weighted NMS of YOLOv5 s post-processing part is changed to DIOU＿NMS.The experiments have supported that the recall rate of the cascade network integrating YOLOv5 s and SRGAN is 1.7 times higher than that of the original YOLOv5 s network model,and the missing rate of small flame is significantly reduced.After mixing attention mechanism,the average detection accuracy of the model was promoted by 11%,indicating the feasibility of introducing attention mechanism.Integrate the above improvement strategies,the detection accuracy reached 98%,the average detection accuracy of flame(IOU=0.5)increased by 44%,and the average detection speed in the test set was 32 frames /s.The trick of pedestrian detection improves the accuracy of target frame regression,and then reduces the error rate of population statistics.