Real-Time Forest Fire Detection by Ensemble Lightweight YOLOX-L and Defogging Method

Area of Science:

• Computer Vision
• Environmental Monitoring
• Artificial Intelligence

Background:

• Traditional forest fire detection methods struggle with fog and rely on computationally intensive features or large models.
• Existing Convolutional Neural Network (CNN) approaches often require extensive parameters, limiting their real-world applicability.
• Fog significantly degrades the performance of automated forest fire detection systems.

Purpose of the Study:

• To develop a lightweight and robust forest fire detection method capable of operating effectively in foggy conditions.
• To improve the accuracy and efficiency of real-time forest fire detection systems.
• To address the limitations of traditional and existing CNN-based fire detection techniques.

Main Methods:

• Proposed GXLD method combines a defogging algorithm (dark channel prior) with a lightweight YOLOX-L model (YOLOX-L-Light).
• YOLOX-L was optimized using GhostNet, depth-separable convolution, and Squeeze-and-Excitation Network (SENet) for reduced parameters.
• Performance evaluated using mean average precision (mAP) for accuracy and network parameters for efficiency.

Main Results:

• The lightweight YOLOX-L-Light model achieved a 92.6% reduction in parameters and a 1.96% increase in mAP.
• The GXLD system demonstrated a mAP of 87.47%, outperforming the original YOLOX-L by 2.46%.
• GXLD achieved an average frame rate of 26.33 fps at 1280x720 resolution, enabling real-time detection in fog.

Conclusions:

• The proposed GXLD method offers a lightweight, accurate, and fog-resilient solution for real-time forest fire detection.
• GXLD's advantages include effective defogging, high target confidence, and target integrity, making it suitable for modern video detection systems.
• This research contributes a practical and efficient system for enhancing forest fire early warning capabilities.