Flame and Smoke Detection Algorithm Based on ODConvBS-YOLO v5s

Abstract

This project  introduces an improved flame and smoke detection algorithm based on YOLOv5s, incorporating the ODConvBS (Ordinary Convolutional Blocks with Spatial Attention) to enhance the extraction of attentional features from the convolutional kernel. Additionally, the model incorporates Gnconv in the Neck to improve high-order spatial information extraction. The traditional algorithms for flame and smoke detection suffer from issues such as low accuracy, high miss rates, low detection efficiency, and poor performance in detecting small targets, leading to significant losses in fire-related incidents. Using a dataset of flame and smoke, the upgraded YOLOv5s model demonstrated a significant improvement,  increase in mean average precision (mAP). The accuracy rate, recall rate, and detection speed also saw substantial enhancements respectively. The proposed algorithm not only addresses the shortcomings of traditional flame and smoke detection methods but also showcases tangible performance improvements, making it a promising solution for real-time and accurate fire detection. The integration of novel components in the model architecture contributes to enhanced feature extraction, leading to better overall detection performance in terms of accuracy, recall, and speed.The project extends its performance capabilities by incorporating advanced detection techniques with YOLO v5x6 and YOLOv8, in which YOLO v5x6 achieved notable metrics of 79.2% mAP, 74% recall, and 80.6% precision.