An Improved Real-Time Detection Method for Flame and Smoke Identification Based on YOLOv5

Fire is a very harmful disaster, which seriously threatens people's life and property safety, so the detection and early warning of fire is very important. With the development of deep learning, smoke, and fire detection technology has become very mature. Despite this, the flame recognition algorithm based on YOLOv5s still has the disadvantages of slow calculation speed, low accuracy, and high false detection and missing rate. To solve this problem, this paper proposes an improved flame and smoke detection algorithm based on the YOLOv5s network, mainly studying how to improve the efficiency and accuracy of flame and smoke detection. First, the fully Attentional network (FLAnet) is introduced to enhance the simultaneous encoding space and channel attention of a single similar graph, and improve the global computing efficiency; Secondly, CoordConv is introduced to make it have space awareness, can carry out identity mapping, and improve the learning rate, and named FC-YOLO. Finally, according to the data set, data enhancement technology and label smoothing are used to preprocess the sample set, which solves the overfitting problem caused by insufficient sample number. The experimental results show that the average mAP@0.5 value is 73.6%, 3.5% higher than the original model. The improved network model has higher detection accuracy, improves the efficiency of fire monitoring, helps to reduce the loss caused by fire, and has great practical significance.