CEVG-RTNet: A real-time architecture for robust forest fire smoke detection in complex environments.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Environmental Monitoring

Background:

• Forest fire smoke detection is critical for early warning systems.
• Complex environmental conditions like low contrast and occlusion challenge existing detection methods.
• Existing systems often struggle with accuracy in real-world, dynamic scenarios.

Purpose of the Study:

• To develop a robust, real-time forest fire smoke detection architecture.
• To enhance detection accuracy under challenging environmental conditions.
• To provide an efficient and effective solution for forest fire early warning.

Main Methods:

• Proposed CEVG-RTNet architecture featuring Spatial-Channel Priori Perceptual Convolution (SCPP-Conv) for improved localization and morphology perception.
• Incorporated Hierarchical Residual Feature Alignment (HRFA) for multi-scale feature extraction and Dynamic Recursive Feature Enhancement (DRFE) for refined dynamic detection.
• Introduced Polygonal-Intersection over Union (PolyIoU) Loss for handling complex smoke morphology and a graph sparse attention mechanism for efficiency.

Main Results:

• CEVG-RTNet-n variant achieved 89.1% precision, 82.9% recall, and mAP@0.5 of 89%.
• Achieved mAP@0.5:0.95 of 58.9% with only 3.04M parameters and 6.6G FLOPs.
• Demonstrated high operational speed with 99.42 FPS, indicating real-time applicability.

Conclusions:

• CEVG-RTNet offers significant improvements in forest fire smoke detection robustness and accuracy.
• The architecture exhibits strong generalization and anti-interference capabilities for complex environments.
• The model's efficiency and performance make it suitable for practical early warning and emergency management systems.