SM-YOLO: A Model for Real-Time Smoke Detection

Abstract

To address the lack of up-to-date smoke detection datasets, we have compiled and labeled a variety smoke detection dataset called SM-dataset. This dataset contains a total of 11596 smoke images from natural scenes. Meanwhile, we introduce a new version of YOLO with better performance, which we call SM-YOLO. SM-YOLO builds on the original model of YOLOv5m, reduces the original three outputs to two, streamlines the original network structure and improves the loss of the original network. Compared with YOLOv5m, SM-YOLO has only 75% of the trainable parameters, but improves mAP@.5 by relative 2%, and reduces the inference time of a single frame from 7.3 ms to 6.6 ms, which effectively improves the speed of smoke detection.