Research On Deep Learning Methods For Forest Fire Detection

Forest fires have occurred frequently in recent years,bringing huge economic losses and casualties to human society,so how to quickly and effectively achieve forest fire detection is a hot issue at present.With the rapid development of deep learning,image recognition and detection are widely used in various fields of life.For the problem of forest fire detection,three forest fire detection methods are proposed in this paper based on the relevant knowledge of deep learning,and the main research contents are as follows.(1)A forest fire detection method based on the channel pruning YOLOv3 algorithm is proposed.The channel pruning idea is introduced into the algorithm to realize the effective compression of the model.The algorithm first performs basic training on the YOLO v3 network to obtain a high-precision algorithm model,then introduces a scaling factor for each network channel,multiplies it with the output of the channel,then performs joint training on the network weights and scaling factors,and finally prunes the channels with smaller scaling factors to fine-tune the pruned network model,and uses a test set to test the performance of the network without sacrificing detection accuracy.The network parameters are reduced to 1/6 of the original size and the forward inference time is changed to 1/2 of the original size.(2)A forest fire detection method based on the MobilenetV3-YOLOv4 algorithm is proposed.The lightweight convolutional neural network Mobilenet and the idea of depth-separable convolution are introduced into the network structure,which effectively reduces the number of network parameters and improves the network detection speed.The MobilenetV1,V2,and V3 networks were used to replace the backbone feature extraction network CSPDarknet-53 of the YOLOV4 network,and the convolution in the YOLOv4 network was replaced with the depth-separable convolution,and the parameters of the replaced networks were compared,and the MobilenetV3-YOLOv4 algorithm based on forest fire detection was finally selected,compared with the Compared with the YOLOv4 algorithm,it sacrifices 1%of detection accuracy to double the detection speed.(3)A forest fire detection method based on CenterNet anchorless frame detection algorithm is proposed.Applying the anchorless frame target detection algorithm to forest fire detection provides a new solution idea.The method transforms the target detection problem into a key point estimation problem by first determining the location of the target’s center point and then returning to the other attributes of the target.The heat map corresponding to the key point is first generated,and the peak of the heat map is the center point of the target,and then the local offset and the width and height information of the bounding box are obtained by regression calculation of the center point,and finally the target bounding box is predicted.The method obtains 84%detection accuracy and 44 FPS detection speed per second on the test set.