Forest Fire Detection Research Based On Deep Learning

Forest is one of the most important resource on the Earth,and plays a vital role in both natural world and human world.However,massive forest fires occurred frequently in recent years and they did a lot of damage to both ecosystem and human society.Thus,it is very important to detect forest fires fast and accurately.The traditional ways of detecting wildfires can not satisfy the need of detecting wildfires nowadays because of many disadvantages such as poor real-time performance and relying on man work.As the deep learning technology developed,many good object detecting algorithm were proposed.However,forest fire detecting has a high demand of detecting speed and detecting accuracy.What’s more,there are lots of interference conditions in forests.So,the exsiting algoritm could hardly meet the requirement.This paper based on deep learning technology,and take research on detecting algorithm of forest fires,the main work is as follows:(1)A self-made dataset contains of about 3000 images with different scales and interference conditions was made in this paper.5 different algorithms were trained and tested on the dataset,and the test results were used to compare with following modified algorithm.(2)In order to reduce the parameters,the backbone network was replaced by Mobile Net V2.Prior box size was optimized and data augmentation algorithm was improved to compensate the detection accuracy loss caused by changing backbone network.After using Mobile Net V2 as backbone network,the detecting speed of the model increased from 97 FPS to 107 FPS.Second,to solve the problem of decreased detecting accuracy,bi K-means cluster algorithm was used to analyse the dataset.The size of prior boxes was optimized according to the results.Lastly,a new data augmentation method,Mask Mosaic,was proposed,and was used to modify the SSD model.The result of test showed that Mask Mosaic has better performance than Mixup and Mosaic.As the experiment showed,the modified SSD model is faster and more accurate than 5 base model.(3)In order to improve detection accuracy of YOLOv7-tiny,attention modules were embeded besides the substitution of activation function.Firstly,introduce 4 attention modules into the YOLOv7-tiny model,which are SE,CBAM,ECA and CA.Test results showed that SE and ECA could improve the detecting accuracy and only slow down the model a little bit,while CBAM and CA significantly slow down the model.Secondly,the activation function was replaced by GELU,Si LU and ACON respectively.In this way,the detecting accuracy was imporved with no detecting speed loss.In the end,the experiment result showed that the combination of SE block and ACON achieved a good balance between speed and accuracy,and introduces a small computational cost.The AP increased by 3.2%,precision increased by 4.44%and recall increased by 3.91%.