Research On Forest Fire Smoke Detection Algorithm Based On Deep Learning

Forest fires have occurred frequently in recent years,threatening the ecosystems of earth and human life seriously.The traditional method of detecting fire-smoke has too many drawbacks and can no longer meet the needs of real-time detection.Since 2012,with the rapid development of convolutional neural networks,target detection algorithms based on deep learning has been continuously proposed and applied to various fields,including forest fire smoke detection.In order to improve the performance of the forest fire smoke detection model and achieve a better balance between detection accuracy and speed,based on the YOLOv4 algorithm,this paper improves the YOLOv4 algorithm and proposes three different forest fire smoke detection methods,and compares the performance advantages of the three algorithms.The details are as follows:（1）An improved YOLOv4 algorithm with embedded attention mechanism was proposed.In order to improve the accuracy of detection,two types of attention mechanisms was embedded before the output of the YOLOv4 algorithm Head: SENet and CBAM.Embedding the attention mechanism in front of the network output has two advantages: firstly,it can reduce the influence of interference information,thereby improving the network’s ability to extract feature information;secondly,the network only added a very small amount of parameters.After embedding SENet,the m AP of the algorithm increased from 93.72% to 94.54%,and the amount of parameters increased 0.33MB;after embedding CBAM,the m AP increased to 96.40%,and the amount of parameters increased by0.7MB.（2）An improved YOLOv4 algorithm based on a lightweight framework was proposed.In order to improve the detection speed,the backbone feature extraction network of YOLOv4 is replaced with lightweight framework Mobilenet network,which has less network parameters,whick can improve the speed of the algorithm to detect forest fire-smoke.The three networks of the Mobilenet series were introduced into the YOLOv4 algorithm for experiments.Among them,the Mobilenetv1-YOLOv4 algorithm had the best detection effect,which increased the detection speed by 60% while only sacrificing 0.78% m AP.（3）A forest fire smoke detection method named Mo Am-YOLOv4（Mobilenet V1 and attention mechanism,Mo Am）was proposed.The purpose is to avoid the decrease of detection accuracy as much as possible while improving the detection speed.This is the integration of the above two contents,which replaces the Backbone of YOLOv4 with the Mobilenet network,and the attention mechanism is embedded before the output of the network,a total of six sets of comparative experiments were obtained.Among them,the Mobilenetv1-YOLOv4+SENet algorithm has the best detection effect,the m AP of this improved algorithm reached 95.16%,and the FPS reached 80.33%..Compared with the original YOLOv4 algorithm,its detection speed FPS was increased by 51% with almost no decrease in m AP.The above three methods are tested and verified on the same dataset,and all use the K-means method to perform cluster analysis on the smoke dateset,so it can obtain candidate boxes that are closer to the smoke images.Experiments have verified the feasibility and effectiveness of several improved algorithms.Compared with the original YOLOv4 algorithm,the three improved algorithms have different improvements.