| Title |
LSA-YOLO: Lightweight Shuffling Attention for Real-time Smoke and Fire Detection |
| Authors |
윤창섭(Chang-Seob Yun) ; 박윤하(Yun-Ha Park) |
| DOI |
https://doi.org/10.5573/ieie.2025.62.8.59 |
| Keywords |
Object detection; Quantization; Multimodal attention; Temporal shift; Small object detection |
| Abstract |
Wildfire-related disasters demand real-time, reliable detection systems capable of handling small-scale smoke and fire under diverse environmental conditions. We propose LSA-YOLO, a novel lightweight object detection framework that integrates Lightweight Shuffling Attention (LSA) to enhance multimodal representation learning and temporal sensitivity. Our architecture introduces a Shuffling CNN backbone for efficient feature mixing, and a MaxReLuCBTS attention mechanism that fuses image and text features to capture subtle visual cues. We further incorporate Flexible InfoNCE Loss for soft text-image alignment and Q-IoU Loss for refined bounding box regression. Evaluations on the D-Fire and HPWREN datasets demonstrate that LSA-YOLO outperforms prior lightweight and fire-specific YOLO variants, achieving state-of-the-art accuracy (up to 60.1% mAP@50?95) while maintaining real-time throughput and minimal latency across HD and UHD resolutions. Our results suggest LSA-YOLO as a competitive model for practical wildfire monitoring and early-warning systems. |