Fire and Smoke Detection Based on Improved YOLOV11

Abstract

Fire and smoke detection plays a vital role in safeguarding human life, infrastructure, and ecological systems by enabling early identification of hazardous events. Conventional vision-based detection techniques rely heavily on handcrafted features, which often fail to deliver reliable performance under complex environmental conditions such as varying illumination, background clutter, and smoke density. To address these limitations, a deep learning–based fire and smoke detection framework is developed using advanced YOLO architectures. A curated image dataset containing fire and smoke scenes is utilized, where bounding box annotations are prepared in YOLO format for object localization. Preprocessing includes image normalization, dataset structuring, and configuration through a unified data file to support both classification and detection pipelines. Multiple YOLO variants, including YOLOv5, YOLOv8, YOLOv9, YOLOv10, YOLOv11s, and an enhanced YOLOv11-DH3 model, are trained and evaluated. In addition, the latest YOLOv11x model is incorporated to assess performance improvements. Model evaluation is conducted using standard metrics such as precision, recall, and mean average precision (mAP). Experimental results demonstrate that YOLOv11x achieves the best performance, attaining a precision of 0.930, recall of 0.981, and mAP of 0.967. The system is further integrated into a Flaskbased web interface to enable real-time fire and smoke detection, enhancing practical applicability and deployment efficiency.