CTM-DETR-frequency-aware and statistically guided transformer for early infrared forest fire detection

Area of Science:

• Computer Vision
• Remote Sensing
• Artificial Intelligence

Background:

• Infrared forest fire detection faces challenges from background thermal interference and weak signals.
• Current spatial-domain methods lack robustness in complex environments, neglecting frequency-domain characteristics.

Purpose of the Study:

• To develop an end-to-end detection framework (CTM-DETR) for robust infrared forest fire monitoring.
• To improve the detection of early-stage forest fires by addressing limitations in existing methods.

Main Methods:

• Introduced a frequency-aware backbone (CGlobalFilter) for real-spectrum filtering in the frequency domain.
• Embedded a statistics-guided linear attention mechanism (TSSA) to reduce interaction complexity from O(N²) to O(N).
• Incorporated a Matching-Aware Loss (MAL) to mitigate sample imbalance through adaptive sample reweighting.

Main Results:

• CTM-DETR achieved a 3.1% mAP50 improvement over RT-DETR on a constructed infrared forest fire dataset.
• The framework demonstrated a 15.6% reduction in parameters and 17.8% lower computational cost.
• The proposed methods effectively suppressed non-fire thermal disturbances and preserved global contextual modeling.

Conclusions:

• CTM-DETR offers superior performance and efficiency for infrared forest fire detection.
• The study provides novel insights into the frequency-domain and statistical properties of infrared flame radiation.
• The proposed framework offers a transferable paradigm for thermal imaging-based perception tasks.