DETR-AE-Mamba: A lightweight and accurate detection framework for early fire warning

Fire disasters pose severe threats to human lives, ecological systems, and economic assets worldwide, making early and accurate detection crucial for timely disaster response. While deep learning-based methods have shown promise for fire monitoring, existing approaches face persistent challenges: edge blurring caused by smoke and adverse weather degrades localization accuracy, small fire sources are frequently missed due to insufficient pixel information, and flame-like backgrounds cause false alarms. Current CNN+Transformer architectures struggle to address these issues simultaneously while maintaining computational efficiency. To bridge this gap, we propose Detection Transformer-Adaptive Encoder-Mamba (DETR-AE-Mamba), a lightweight detection framework with three novel components: Mamba-Perceptron (Mamba-P), which processes features in both time and frequency domains to enhance edge extraction; Adaptive Fusion Encoder (AFE), which improves global perception for small object detection; and ELD-Decay-Varifocal-Loss, a dynamic loss function that enhances noise robustness. Experiments on the D-Fire dataset show that DETR-AE-Mamba improves precision by 3.4% over RT-DETR-R18 while reducing parameters by 54.2% (from 21.4M to 9.8M) and GFLOPs by 69.8% (from 56.9 to 17.2), demonstrating potential for fire detection and applications in other fields.