A lightweight fire detection algorithm for small targets based on YOLOv5s.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Fire Safety Engineering

Background:

• Current fire detection algorithms struggle with low accuracy and poor recognition of small targets in complex settings.
• Existing methods often lack the efficiency required for real-time applications.

Purpose of the Study:

• To develop a lightweight and accurate fire detection algorithm.
• To improve the recognition of small fire targets in challenging environments.
• To meet the real-time detection demands for enhanced fire safety.

Main Methods:

• An improved YOLOv5s architecture incorporating the Contextual Transformer (CoT) and a novel CSP1_CoT module.
• Enhancements to the Neck architecture with a dedicated small target detection layer and SE attention mechanism.
• Implementation of the Focal-Efficient IoU (Focal-EIoU) loss function for improved convergence and precision.

Main Results:

• The modified model achieved a mean Average Precision (mAP@.5) of 96% and an accuracy of 94.8%, an 8.8% and 8.9% improvement, respectively.
• Reduced model parameter count by 1.1% to a compact 14.6MB size.
• Achieved a detection speed of 85 Frames Per Second (FPS), meeting real-time requirements.

Conclusions:

• The enhanced YOLOv5s algorithm offers superior precision and accuracy for fire detection.
• The lightweight design and high detection speed satisfy real-time and resource-constrained application needs.
• This improved algorithm effectively addresses the limitations of current fire detection systems.