Improving YOLOv5s Algorithm for Detecting Flame and Smoke

Abstract

Object detection methods can be used to detect flames and smoke from images or videos for the identification and exploration of fire. In this paper, an improved YOLOv5s algorithm, called GAM-ASFF-YOLOv5s is proposed, which introduces an attention mechanism and feature fusion layer. A global attention mechanism (GAM) is introduced into the backbone network of YOLOv5s to focus on the detection information that is conducive to flames and smoke, while suppressing unimportant information. A head network with adaptive spatial feature fusion (ASFF) was designed to extract more complete image features of flame and smoke. In addition, the original object bounding box regression loss function the complete IoU (CIoU) of YOLOv5s was replaced by repulsion loss to enhance the generalization ability of the model and further improve the detection performance of the flame and smoke. The experimental results show that the precision of the GAM-ASFF-YOLOv5s+REP algorithm was 5.7% higher than that of the original YOLOv5s algorithm on the VOC2007 dataset and it also performed well on the flame and smoke dataset, that is, the precision, recall, and mean average precision (mAP) were improved by 0.5 %, 2.1 % and 2.9 % respectively, compared to the original YOLOv5s algorithm.

Keywords

• Adaptive spatial feature fusion
• flame and smoke detection
• global attention mechanism
• repulsion loss
• yolov5s