Research On Method For Forest Fire Detection Based On Deep Learning

Forest resources are one of the important resources for human survival.Forest fires not only cause huge damage to forest resources,but also bring serious losses to humans.Therefore,the real-time and accurate forest fire detection is significant to protection of forest resources.Forest fire detection is an application scene of target detection,which includes two tasks of fire identification and fire location.That is to identify the fire in the image and find location of the fire and mark it out.Fire detection includes smoke detection and flame detection.Smoke occurs earlier in the forest fire than flame.It can be used as an important sign for early detection of forest fires.In order to detect forest fire timely,smoke is selected as the object of forest fire detection.The typical one stage detection model YOLO v3 and two stage detection model Faster R-CNN are analyzed.The experimental comparison and analysis of the models on forest fire smoke detection are carried out.Through the research and comparative experiments,the YOLO v3 algorithm is selected as the basic model of forest fire smoke detection.In order to solve the problems of low local feature utilization and insufficient multi-scale feature fusion in YOLO v3 algorithm,the spatial pyramid pooling network and path aggregation network are deeply studied.The spatial pyramid pooling module and feature pyramid path aggregation module are constructed based on the YOLO v3 algorithm and two modules are embedded into the YOLO v3 algorithm to form the YOLO v3-SPP-PAN model.The spatial pyramid pooling module can extract local and global features effectively.The feature pyramid path aggregation module can effectively utilize the low-level location features and fuse multi-scale feature.The precision,recall and mean average precision of the optimized model have been improved and are above97%.The detection confidence of YOLO v3-SPP-PAN reaches 93% in the actual forest fire scene without fog.For the problems of slow convergence of YOLO v3-SPP-PAN model and low confidence in fog detection scene,the attention module of CBAM is researched deeply and embedded into the YOLO v3-SPP-PAN model to form CBAM-YOLO v3-SPP-PAN model.The CBAM module can focus on the feature of forest fire smoke through channel attention and spatial attention.The convergence of the optimized model is accelerated and the test precision has increased by 0.4%.In the actual fire scene with fog,the detection confidence of the optimized model has increased by 8%.The experimental results show that the CBAM-YOLO v3-SPP-PAN model can effectively detect forest fire.