Flame and Smoke Detection Based on Improved YOLOv5

Abstract

To provide real-time fire warning based on images，an improved flame and smoke detection algorithm based on YOLOv5s is proposed in this study.The original path aggregation network module is replaced in the Neck of YOLOv5s with a bidirectional cross-scale fusion module，so that the deeper network can directly extract superficial features，enhance information flow，and improve network feature fusion capabilities.The reasoning layer for Coordinate Attention（CA） is added to the Head of YOLOv5s to enhance the ability of extracting and locating network information and improve detection accuracy without increasing the computational burden extensively.A variety of data enhancement technologies，such as Hue，Saturation，Value（HSV） color gamut enhancement，random rotation，and Mosaic，are used to adjust and expand the training data.The k-means clustering algorithm is used to obtain the prior anchor of the dataset to enhance the model robustness.The experimental results demonstrate that compared with the flame and smoke detection algorithm based on YOLOv5s，the mean Average Precision（mAP） of the improved algorithm has increased by 3.2 percentage points，while the detection speed has reached 243 frame/s.The lightweight advantage of YOLOv5s has also been maintained.This is significant for flame and smoke detection in sheltered，dark，small，and multiple targets.