Early-stage infrared forest fire detection is severely hindered by strong background thermal interference and extremely weak fire radiation signals. Existing methods mainly rely on spatial-domain modeling and overlook the frequency-domain characteristics of flame thermal radiation, limiting robustness in complex environments. To address this challenge, we propose CTM-DETR, an end-to-end detection framework tailored for infrared forest fire monitoring. A frequency-aware backbone, CGlobalFilter, is introduced to explicitly model thermal radiation priors by performing real-spectrum filtering in the frequency domain, effectively suppressing non-fire thermal disturbances. Furthermore, a statistics-guided linear attention mechanism (TSSA) is embedded into the detection head of AIFI. This mechanism approximates the dominant global interactions in conventional pairwise attention using token-level second-order statistics, thereby reducing the interaction complexity from O(N²) to O(N) while preserving global contextual modeling ability. To mitigate sample imbalance, a Matching-Aware Loss (MAL) is incorporated to adaptively reweight samples based on matching quality. Experiments on a constructed infrared forest fire dataset show that CTM-DETR surpasses RT-DETR, achieving a 3.1% mAP50 improvement, with 15.6% fewer parameters and 17.8% lower computational cost. Beyond performance gains, this work provides new insights into the frequency-domain and statistical properties of infrared flame radiation and offers a transferable paradigm for thermal imaging-based perception tasks.
Mu et al. (2026) studied this question.