Abstract

Real-time and accurate detection of flame and smoke is an important prerequisite to reduce the loss caused by fire. There exists some problems in traditional flame and smoke detection algorithm, such as low accuracy, high miss rate, low detection efficiency and low detection rate of small targets. This paper proposes a YOLOv5s flame smoke detection algorithm based on ODConvBS. Firstly, in the YOLOv5s backbone network, the ordinary convolution block is replaced by ODConvBS to realize the extraction of attention features of the convolution kernel; Secondly, Gnconv is introduced into Neck to improve the model’s high-order spatial information extraction ability; then the Shuffle Attention module is added at the end of Neck to facilitate the fusion of different groups of features; At last, in the prediction section, the SIOU loss function, which can account for the angle of the prediction frame vector, is utilized to speed up model convergence. When utilizing the self-made flame and smoke data set, the upgraded YOLOv5s model mAP grew by 9.3%.At the same time, the accuracy rate and the recall rate and the detection speed increased to 83.5%, 83.7%, 33.3FPS respectively.

Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call