Real-Time Smoke Detection Algorithm with Transformer

Smoke detection technology plays a crucial role in fire prevention and is vital for ensuring the safety of people's lives and property. However, existing smoke detection models face challenges such as false positives in complex environments and missed edge detections of smoke, especially for small smoke targets. Additionally, high computational and inference costs limit their application. To address these challenges, this paper proposes a Real-Time Smoke Detection Algorithm with Transformer, named RT-DETR-Smoke. First, to reduce computational costs, we introduce a real-time, end-to-end object detector, which includes a high-efficiency hybrid encoder and uncertainty minimization in query selection. Next, to address the issue of missed smoke edge detection, we employ Coordinate Attention to facilitate the fusion of smoke features across different spatial locations. Finally, to accommodate the multi-scale characteristics of fire and the complexity of various environments, we propose the WShapeIoU loss function to accelerate model convergence and improve smoke detection accuracy. Experimental results on a custom smoke dataset show that RT-DETR-Smoke achieves an accuracy of 87.75% mAP@0.5 and excels in real-time performance, with a processing speed of 445.50 FPS, demonstrating its potential for practical applications. Compared to existing models, the proposed method significantly improves both accuracy and speed.