The Research On Early Forest Fire Detection Algorithm Based On Deep Learning

In recent years,forest fires have occurred frequently all over the world,seriously threatening the earth’s ecosystem and human life.If high precision real-time detection can be carried out in the initial stage of forest fire to prevent the spread of forest fire in time,it can reduce the difficulty of fighting and avoid causing greater losses.Aiming at the problems existing in the current early forest fire detection algorithm,such as single forest fire target,low temperature ignition point,blocking target,small target,thin diffusion smoke detection ability,poor generalization and robustness,based on the deep learning technology,the target detection algorithm is screened and improved,and a high-precision,low-consumption,easy to deploy,easy to train early forest fire and small target forest fire detection model is constructed.In order to complete the multi-class,multi-scale and multi-target real-time forest fire detection task in the complex forest environment.The main work of this paper is as follows:(1)Early forest fire data set and early forest fire small target data set are established.According to the research requirements,a complex and changeable image data set with the characteristics of early forest fire covering different regions,different seasons,different light and different scale features was established.After data enhancement,data set annotation and format standardization,early forest fire data set and early forest fire small target data set could be used for experiments.(2)A target detection algorithm that meets the detection requirements is constructed and screened based on deep learning.Five early forest fire detection models were constructed based on Faster R-CNN,CenterNet,SSD,YOLOV4 and YOLOV5s target detection algorithms.Then,a comprehensive evaluation was conducted on YOLOV5s from six aspects:training process,detection accuracy and miss rate,real-time and lightweight,and actual detection ability.The experimental results show that the mAP value of Yolov5S is up to 91.91%,the average miss rate is 18.5%,the FPS is 72 frames/s,and the anti-interference ability is strong.Can better cope with the early forest fire detection task.(3)An improved YOLOV5s small target detection algorithm for early forest fires is proposed.Based on YOLOv5s algorithm,improved F-RFB enhanced receptive field module is added to Backbone and SPP module is replaced by SimSPPF module,which improves the feature extraction capability and context information capture capability of network.ECA attention module is added to Neck to improve the feature fusion capability of the network.Focal Loss function and transfer learning were introduced in the training process to improve the training efficiency.The experimental results show that the mAP values of the improved YOLOv5s algorithm on the two data sets are 94.90%and 90.66%respectively.The average missed detection rate was 13.5%and 24.5%,respectively.Compared with before improvement,mAP values increased by 2.99%and 10.01%,respectively,and the average missed detection rate decreased by 5%and 15.5%,respectively.The results show that the performance of the improved model is improved effectively and has a broad application prospect.