DSS-YOLO: an improved lightweight real-time fire detection model based on YOLOv8

Area of Science:

• Computer Vision
• Artificial Intelligence
• Fire Safety Engineering

Background:

• Fire disasters present significant risks, with early detection challenged by small flames, diffuse smoke, and obstructed objects.
• Existing systems often struggle with missed detections and poor real-time performance in complex fire scenarios.

Purpose of the Study:

• To develop an improved fire detection model (DSS-YOLOv8) based on YOLOv8n.
• Enhance recognition accuracy for obscured objects and small targets.
• Reduce computational overhead for real-time applications.

Main Methods:

• Modified YOLOv8n architecture, replacing C2f modules with DynamicConv for reduced computation.
• Incorporated SEAM attention mechanism to improve detection of obscured and small targets.
• Integrated SPPELAN module to enhance multi-scale detection capabilities.

Main Results:

• DSS-YOLOv8 achieved a 0.6% increase in mAP and a 1.6% increase in Recall compared to original YOLOv8n.
• Model size and FLOPs were reduced by 3.4% and 12.3%, respectively.
• Demonstrated improved performance on diverse fire scenarios (indoor, forest, building).

Conclusions:

• The DSS-YOLOv8 model offers effective technical support for intelligent fire monitoring systems.
• Significantly reduces computational cost, enhancing real-time fire detection capabilities.
• Facilitates early detection of fire hazards, minimizing potential damage.