Smoke Of Dilated Depth Normalized Convolutional Neural Network In Campus Environment

Fire is considered to be one of the most serious natural disasters in the world.If we can accurately and promptly respond to each fire situation,it will greatly affect our daily life.Therefore,this article regards fire detection as an important Subject to study.Fire detection methods are mainly divided into fire detection based on smoke information and fire detection based on flame information.This article mainly focuses on the algorithm research of smoke,and proposes an improved detection algorithm for smoke.At present,most of the detection algorithms for smoke can obtain relatively good detection results,but there are still some problems.For example,when the algorithm is just before the fire,the smoke is relatively thin and the moving speed is relatively slow,and the final detection result of the algorithm is not ideal.Most of the original smoke detection methods use initial pictures or hand-synthesized features directly as the input unit of the neural network.The final algorithm has relatively poor robustness,and the smoke scenes are different.The original algorithm is not the same.Suitable for all smoke scenes.This paper proposes an improved fire detection method,the main work is as follows:(1)This paper proposes an improved background difference algorithm,combined with dark channel prior theory to extract candidate smoke.First,the improved background difference method is used to calculate the foreground image.In the past,the moving speed of the smoke was relatively slow,which would bring the problem of holes in the detection results.The improved algorithm solved this problem.Then introduce a dark channel prior method for interfering objects.The traditional algorithm is not very adaptable to the complex smog environment.If there is a thunderstorm in the experimental environment,the wind will blow the branches or many people on the road quickly.Moving,these interference situations can cause false detections.The dark channel prior method combined with the improved background difference algorithm can filter out most of the interference data;finally,feature extraction and feature fusion are performed to achieve the extraction of candidate smoke.The results of the experiment prove that the improved background difference algorithm combined with the dark channel prior can reduce the false detection rate,eliminate holes,and improve the detection performance of the algorithm.(2)This paper proposes an improved convolutional neural network,The network is composed of 16 layers.Its function is to reduce the training time.When the data set is too small or the number of smoke images and non-smoke images is large,it will cause overfitting.Therefore,we use a data enhancement method.Experimental results show that the improved method has a very low false alarm rate and improves the accuracy of smoke detection.