Enhanced Fire and Smoke Detection Using Advanced YOLOv11 Variants for Improved Accuracy

Abstract

Fire and smoke detection is essential for protecting human life, infrastructure, and ecosystems by facilitating the early recognition of dangerous incidents. Traditional vision-based detection methods depend significantly on manually produced features, which frequently do not provide consistent performance in diverse environmental situations, like fluctuating illumination, background clutter, and smoke density. A deep learningbased system for fire and smoke detection is built utilizing enhanced YOLO architectures to mitigate these constraints. A curated image collection featuring fire and smoke scenarios is employed, with bounding box annotations formatted in YOLO for object localization. Preprocessing encompasses picture normalization, dataset organization, and configuration via a consolidated data file to facilitate both classification and detection workflows. Various YOLO variations, such as YOLOv5, YOLOv8, YOLOv9, YOLOv10, YOLOv11s, and an improved YOLOv11-DH3 model, are subjected to training and evaluation. The latest YOLOv11x model is utilized to evaluate performance enhancements. Model evaluation is performed with conventional metrics including precision, recall, and mean average precision (mAP). Experimental findings indicate that YOLOv11x exhibits superior performance, achieving a precision of 0.930, a recall of 0.981, and a mean Average Precision (mAP) of 0.967. The system is additionally included into a Flask-based web interface to facilitate real-time fire and smoke detection, improving practical applicability and deployment efficiency.