Dynamic Smoke Detection Based On Convolutional Neural Network

Every year,many wild fires will occur all over the world,causing great damage to human life and property and the natural environment.Smoke is the most prominent feature in the early stages of fire onset.Therefore,smoke detection can play an important role in early fire warning.With the popularization of deep learning and video surveillance,the research direction of wild fire alarm has gradually changed to dynamic smoke detection based on deep learning.Under the condition of ensuring the real-time detection,the dynamic smoke detection method should improve the alarm accuracy and reduce the missed detection rate as much as possible.Since there is no mature and unified data set for dynamic smoke detection,this paper organizes the data set collected from various aspects to obtain the data set for the experiment.In order to avoid the recognition and processing of the whole video image,based on the sparse optical flow,this paper designs an optimization algorithm for the extraction of suspected smoke areas based on the motion and color characteristics of smoke.Firstly,sparse optical flow is used to obtain feature points with optical flow vectors.Secondly,the optical flow component in the vertical direction of the feature point is used to judge whether the feature point is upward,and the image block with the size of 32*24 is segmented with the upward point as the center.Finally,in order to further improve the possibility that the image to be segmented contains smoke,this paper studies the color characteristics of the smoke image,determines that the smoke has significant characteristics in saturation,and sets a threshold to screen the image blocks to be recognized.This algorithm can effectively improve the accuracy of recognition and reduce the workload of recognition.In order to accurately classify suspected smoke image blocks,this paper determines Resnet34 as the basic architecture of recognition model,and trains and optimizes the model.Finally,the video data set is used to test the detection method designed in this paper,and it is compared with several control methods.The experimental results confirm that the detection method designed in this paper can have better detection effect on smoke video images.Figure [33] Table [12] Reference [80]...