FSD-DETR: A Lightweight Fire and Smoke Detection Model Based on Improved RT-DETR

To address the computational efficiency challenges of fire and smoke detection, we present FSD-DETR, an en-hanced model that effectively resolves the accuracy-efficiency trade-off in complex environments. Our approach incorporates the StarNet backbone to streamline the architecture while reducing computational requirements. We integrate the HiLo attention mechanism into the AIFI module for multi-scale feature interactions, replace RepC3 blocks with gConv modules for enhanced nonlinear representation, and employ WTConv layers for improved multi-scale feature extraction. Comprehensive evaluation demonstrates that FSD-DETR achieves 5.9 percentage points mAP50 improvement on the DFS dataset while reducing parameters by 51.4% and FLOPs by 58.2%. Compared to the RT-DETR baseline, FSD-DETR achieves superior accuracy with substantially reduced computational requirements, establishing an optimal balance between efficiency and detection performance.