| As a disaster with high frequency,fire spreads quickly and is highly destructive.With the widespread popularity of video monitoring,the automatic detection and early warning of fire based on image and video is of great significance and value.Traditional flame detection usually designs feature extraction algorithms based on prior knowledge,extracts dynamic or static features of flame manually,and then performs flame recognition.This type of traditional flame detection method based on computer vision is difficult and time-consuming to design the artificial feature extraction algorithm according to the prior knowledge.When facing different complex environments and variable flame types,its generalization ability is often insufficient.With the rapid development of computer vision based on deep learning in recent years,convolutional neural network can automatically learn and acquire image features,and the object detection technology has a great development and innovation.Aiming at the problem of fire video detection in complex environment,this paper proposes a RetinaNet flame image detection algorithm combined with attention mechanism.On top of a feedforward ResNet architecture,RetinaNet uses the feature pyramid network as the backbone network to generate a rich convolution feature pyramid.The output of the backbone network is classified and regressed separately by the classification subnet and box subnet.Attention mechanism enables the network to locate the information of interest more effectively,and improves the feature expression ability of network model.The improved algorithm has a flame detection rate of 4.5 percentage points higher than the original RetinaNet algorithm.It realizes real-time and end-to-end flame recognition and location,and also maintains a high detection accuracy for small fire point detection under complex object background.For the detection of fire video stream,a CornerNet-Squeeze flame video detection algorithm based on key points is proposed.This algorithm uses hourglass network to extract the flame features,combines the corner pool to detect the pair of corners of the flame to locate the flame.The improved Fire module is used to replace the residual block,which greatly improves the detection rate while losing only a small accuracy.Compared with the traditional flame detection method and the YOLOv3 flame detection algorithm,this algorithm has better detection accuracy and speed.It basically meets the needs of real-time fire video detection and analysis. |
