Abstract:
Fire and smoke detection is essential in safety-critical environments, yet traditional systems
often struggle with maintaining accuracy and reducing false alarms in complex scenarios. Therefore,
vision-based systems are used for preventing fire tragedies. There are different machines and deep
learning techniques used to timely and effectively detect the fire/smoke and one of them is “You Only
Look Once” (Yolo). Yolo is a type of neural network (CNN), which is good at detecting patterns in
images. Yolov8 is the most widely used object detection model for vision-based systems. However,
there still exist some challenges, such as high computational complexity and low detection
performance. This study introduces a novel lightweight and optimal Yolov8 model to over these
challenges. To enhance performance, Efficient Channel Attention (ECA) is integrated into the
model’s head to focus on critical features, while the C3Ghost module in the backbone reduces
computational overhead without sacrificing accuracy. The model is trained and evaluated on two
datasets: FS and FASDD comprising diverse indoor, outdoor fire and smoke scenarios and has
achieved a mAP@50 of 89%, precision:88%, recall: 84%, and an F1-score of 86.4% which shows an
improvement of 4.56% in precision, 2% of recall and 8.10% in mAP@50 in comparison with the
existing state of the art. Our findings have demonstrated significant improvements in detection
accuracy and false-positive reduction compared to other computationally intensive models like
Yolov5, Yolov7 and (vision) Transformers. Our model is lightweight architecture, more accurate in
fire, and smoke detection, and makes it suitable for embedded device deployment.
