Improved Real-Time Smoke Detection Model Based on RT-DETR

Fire remains a major threat to society and economic activities. Given the real-time demands of smoke detection, most research in deep learning has focused on Convolutional Neural Networks. The Real-Time Detection Transformer (RT-DETR) introduces a promising alternative for this task. This paper extends RT-DETR to address challenges such as morphological variations and interference in smoke detection by proposing the Realtime Smoke Detection Transformer (RS-DETR). RS-DETR uses smoke images with concentration data as input and employs a deformable attention module to manage morphological changes, enabling robust feature extraction. Additionally, a Cross-Scale Smoke Feature Fusion Module (CS-SFFM) is integrated to enhance detection accuracy for small and thin smoke targets through multi-scale feature resampling and fusion. To improve convergence speed and stability, Efficient Intersection over Union (EIoU) replaces Generalized Intersection over Union (GIoU) in feature scoring. The improved model achieves an average precision of 93.9% on a custom dataset, representing a 5.7% improvement over the original model, and demonstrates excellent performance across various detection scenarios.