Research On Forest Fire Detection Method Based On Deep Learning

Forest is a valuable treasure of human beings,however,a forest fire is easily to destroy the forest,therefore research on forest fire detection technology has important practical significance.Due to its high detection accuracy and fast detection speed,the deep learning-based target detection algorithm is increasingly used in forest fire detection in combination with UAV.In this context,the research on forest fire detection methods based on deep learning has become a hot topic for researchers in relevant fields.Forest fire detection based on deep learning by UAV camera to inspection of the forest,by wireless transmission system will be checking video background or video directly into the onboard data processing platform,through the use of deep learning algorithm to back to the video in real time detection,to detect whether there is a fire in the video,smoke or smoke plus fire.In order to solve this problem,this thesis mainly carries out the following works:1)on the ground end computer based on Gaussian YOLOv3 forest fire detection algorithms,this algorithm using Gaussian model to the modeling of Bounding Box,in the basic don't change YOLOv3 structure and amount of calculation,reliability can be estimated output per box,and redesigning the loss function,improve overall performance of detection algorithm in ascension,at the same time in the process of detection using positioning uncertainty of prediction,can significantly drop FP(false positives),improve the accuracy(TP);2)an improved PANet forest fire detection algorithm was proposed for the computer at the ground end.This algorithm enhanced the whole feature level with accurate positioning signals at a lower level through bottom-up path expansion,thus shortening the information path between the bottom and top features.The adaptive feature pool is used to connect the feature grid with all feature layers,so that the useful information in each feature layer can be directly transmitted to the following suggested subnet.While improving the backbone network with CSPNet,GIOU loss function is introduced to improve the detection accuracy;3)due to the weak computing capacity of the UAV airborne processor,the network with fewer layers can only be selected.Therefore,the improved forest fire detection algorithm of SSDLite is proposed,which introduces the reverse residual of the linear bottleneck.The low-dimensional compressed representation is used as the input,which is first extended to a higher dimension,and then filtered by a lightweight depth convolution.The features are then projected back to the lower dimension by linear convolution.At the same time,the improved CSPNet backbone network was introduced to reduce the amount of operation and memory required while maintaining the same accuracy.4)at present,there is no public forest fire data set.Through downloading forest fire pictures from the Internet and framing forest fire videos,and manual screening of all pictures,and 8000 pictures were selected and manually labeled with LabelImge to obtain a forest fire data set containing flame,smoke,smoke and fire for subsequent experiments.Through a large number of comparative experiments,the forest fire detection method proposed in this paper has achieved good results.The final experimental results show that the gauss-YOLOv3 recognition rate is 94%and the detection speed is 47fps on the ground server,the improved PANet recognition rate is 91%and the detection speed is 74.2fps,and the improved SSDLite recognition rate is 85%and the detection speed is 6.2fps on the UVA airborne processor.