Research And Implementation Of Fire Detection Algorithm Based On Deep Learning

Fires are a serious threat to the safety of people's lives and property,so early fire detection research is of great significance and value.Traditional fire detectors are mainly physical sensors that detect the occurrence of fire by detecting changes in smoke,light,and gas.These detectors generally have problems of small coverage and single use scenarios.In recent years,Visual fire detectors using monitoring images have received increasing attention.At the same time,with the deepening of deep learning image processing technology,the deep learning algorithm can extract the smoke and flame features in the image better than the traditional manual feature extraction operator,and the detection accuracy and speed are greatly improved.Based on the deep learning algorithm,the fire image detection is studied.This paper proposes a fire detection scheme jointly implemented by smoke detection and flame detection.In the smoke detection algorithm design,this paper replaces the original backbone network of YOLO V3 with Densenet121,which constitutes the YOLO-Densebackbone network,and performs online image enhancement operations on the self-built training set.Flame detection firstly adopts a scheme based on pre-selected regions and classifiers.The algorithm filters the suspected flame regions by the motion characteristics and color features of the flame,and fills the holes through the secondary color region expansion algorithm.The inter-frame continuity feature of the bounding box and the color pixel area ratio are further filtered to select the pre-selected areas that meet the conditions.Finally,the flame pre-selected area image is classified by the PCB network.A second scheme using the anchor-free Center Net for flame recognition is proposed.The network is based on the flame center point for prediction.By transforming the sample set label into a characteristic heat map for network training,the DLA-34 network is selected on the backbone network and regression is performed using Focal Loss.Finally,the test results show that the YOLO-Densebackbone network has an AP value of 87.93 in the self-built test set for smoke detection,and the parameter is reduced by half.The flame detection algorithm based on the preselected area and classifier and the flame detection algorithm based on Center Net can detect the flame within 5s,and the ratio of the detected flame frames to the total frame number is greater than 80%.