Research On Forest Fire Image Detection Method Based On YOLO

Forest fires are sudden and destructive,they will cause significant damage to both human beings and the natural environment once they happen.The ability to quickly and accurately detect forest fireworks has become the key to protecting forests and stopping losses in time.At present,many forest fire detections are based on fire detectors and sensors but the alarm rate is high,the visualization is not enough,and the detection delay is large.It is difficult to accurately identify and return fire information in the early fire.In recent years,the use of deep learning-based technology for target detection has become increasingly popular,and it has certain practicability in the forest firework detection scene.This topic is based on the deep learning YOLO model detection algorithm,and proposes and improves two detections’ method based on different application scenarios,the specific work is as follows:(1)Construction of the forest fire dataset.Currently,there are few forest fire datasets.This topic constructs a dataset suitable for forest fire detection laying the foundation for subsequent algorithm training through network search and screening.(2)A lightweight network improvement based on YOLOv5.In order to achieve forest fireworks detection on mobile devices such as embedded systems,an improved lightweight YOLOv5 algorithm is proposed to meet the characteristics of small memory and low computing power of mobile devices.The algorithm reduces the size of the network model parameters while improving the detection accuracy,making it suitable for forest fireworks detection with good detection performance while meeting the requirements of lightweight design.Firstly,the lightweight Mobile Netv3 network is used at backbone network to improve the YOLOv5.The introduction of adaptive width,inverse residual structure,and SE attention mechanism module makes the network have better feature extraction capability and effectively reduces the number of parameters and computation in the network,improving the detection efficiency.The use of the weighted bidirectional feature pyramid network Bi FPN enables better cross-scale feature fusion of the feature maps extracted by the network,improving the detection accuracy of the lightweight network.Compared with the baseline YOLOv5,this improved algorithm has certain improvements in the detection accuracy and speed of forest fireworks,which verifies the effectiveness and feasibility of the algorithm.(3)Improvement of detection algorithm for fixed application scenarios.In detection scenarios where GPU computing power is abundant,this topic proposes and improves a forest fireworks detection algorithm based on the YOLOv5 model to enhance detection accuracy.Firstly,the basic location anchor boxes used for training are obtained using the k-means clustering algorithm based on 1-Io U on the self-built forest fireworks dataset.Transformer attention mechanism is introduced for subsequent attention calculation and weight assignment,allowing the model to better understand and capture key features when processing sequential data,thereby improving the performance of the model.The improved bounding box loss function,EIo U,is used to evaluate the matching degree between the predicted and true bounding boxes,with differences in width and height being calculated to more accurately reflect the overlap between the two bounding boxes.Finally,in order to reduce network parameters and improve the speed of algorithm network detection,some convolutional layers are improved to depth-wise separable convolution.The m AP of the improved algorithm reached 0.929,and the FPS value reached 76.92,meeting the requirements for forest fireworks detection.