Research On Forest Fire Image Warning Detection Method Based On Anchor-Free Network

Forest fire detection is an important component of forest fire monitoring and early warning system,which is of great significance for protecting forest resources,preventing fire accidents,and maintaining social security.Traditional methods usually use image processing technology and manually extracted features to detect fires,but this method requires manual intervention and has limited effectiveness in complex scenes.With the continuous development of deep learning technology,some object detection models suitable for forest fire detection have emerged,but most of these models are based on anchor points,requiring manually set anchor boxes.If the number and size of anchor boxes are not set reasonably,it may result in incomplete coverage of the target or increased computation.In addition,when the size and shape differences between the target and anchor boxes are large,misjudgments or omissions may occur,which is very unfavorable for smoke and fireworks detection with varying sizes in forest fires.Therefore,this dissertation adopts an anchor-free deep learning object detection algorithm to achieve real-time detection for forest fires.The specific work is as follows:(1)Construct forest fire data set.One of the challenges in this research is the lack of publicly available standard forest fire datasets.To address this issue,this thesis collects and screens forest fire images,including Baidu Image Library,Paddle dataset,and FLAME dataset,etc.The collected images were then cleaned to remove low-quality and irrelevant images,and data augmentation was applied to increase the quantity and diversity of the data.Finally,the images were annotated to build a dataset mainly focused on forest fire scenes.(2)An improved forest fire detection algorithm based on anchor-free CenterNet is proposed for low-computing devices such as drones.Firstly,the algorithm adopts a lightweight and efficient backbone network,Vo VNet,and introduces the CSP structure to reduce computational complexity and improve detection speed.Then the feature fusion network FPN is used to combine feature maps from different scales,improving the detection accuracy and positioning effect of early-stage small smoke fires.Finally,the loss function is optimized to make the model predict more accurately,reducing missed detections and false alarms.(3)An improved forest fire detection algorithm based on two-stage CenterNet2 is proposed for high-computing devices such as ground servers.The improved CenterNet with optimized loss function is used as the first stage detector,further improving detection accuracy.Vo VNet V2 is used as the backbone network,and the asymmetric convolution is introduced to increase the size of the receptive field by combining the vertical and horizontal direction convolution kernels.The bidirectional weighted connection structure and e SE attention mechanism are introduced on the basis of FPN,making the network more accurate in expression.For the regression loss in the second stage,the SIo U loss function is used,introducing the cost of the direction angle between the true box and the predicted box to improve the detection accuracy of early-stage fires,and the function is simplified to improve the stability and efficiency of training.The experimental results demonstrate that the forest fire detection method proposed in this thesis can effectively improve the accuracy of smoke and fire detection in tree-covered scenes,inter-forest lighting scenes,and snowy environments,and reduce the occurrence of missed alarms and false alarms.In addition,it can effectively improve the accuracy of detecting small smoke or small fireworks captured from a distant shooting distance.The improved CenterNet achieves a recognition rate of 86.4% and a detection speed of 51.2frames/second,while the improved CenterNet2 achieves a recognition rate of 90.6% and a detection speed of 36.8 frames/second.