Scholarlink

<p>To overcome the limitations of conventional fire detection methods in accurately recognizing small-scale, multi-target flames and irregularly shaped, low-quantity smoke due to insufficient feature extraction, this study proposes an enhanced YOLOv8-based algorithm called FG-YOLO. By integrating the FSCNet backbone and the GScELAN efficient layer aggregation module, FG-YOLO effectively mitigates feature detail loss that commonly arises from convolutional iterations. This advancement significantly boosts the representation ability for multi-scale, occluded, and small-object features, thereby improving detection accuracy while reducing computational overhead. The proposed approach was evaluated on the FASD dataset, which comprises 15,778 manually annotated images of flames and smoke labeled with the LabelImg tool. Experimental results show that FG-YOLO achieves a 5.4% increase in accuracy, a 6.8% increase in recall, a 4.5% increase in mAP50, and a 9.9% increase in mAP50-95 compared to the YOLOv8 baseline, alongside an inference speed of 76.51 FPS. With only 2.8 million parameters, FG-YOLO delivers high performance with optimized computational efficiency, offering a robust solution for real-time fire and smoke detection in safety-critical applications.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

FG-YOLO: an improved YOLOv8 algorithm for real-time fire and smoke detection

  • 期刊论文

  • 发表日期：

    2025年3月1日

查看全文
  • Jiale Yao,
  • Juyang Lei,
  • Jun Zhou,
  • Chaofeng Liu

摘要

To overcome the limitations of conventional fire detection methods in accurately recognizing small-scale, multi-target flames and irregularly shaped, low-quantity smoke due to insufficient feature extraction, this study proposes an enhanced YOLOv8-based algorithm called FG-YOLO. By integrating the FSCNet backbone and the GScELAN efficient layer aggregation module, FG-YOLO effectively mitigates feature detail loss that commonly arises from convolutional iterations. This advancement significantly boosts the representation ability for multi-scale, occluded, and small-object features, thereby improving detection accuracy while reducing computational overhead. The proposed approach was evaluated on the FASD dataset, which comprises 15,778 manually annotated images of flames and smoke labeled with the LabelImg tool. Experimental results show that FG-YOLO achieves a 5.4% increase in accuracy, a 6.8% increase in recall, a 4.5% increase in mAP50, and a 9.9% increase in mAP50-95 compared to the YOLOv8 baseline, alongside an inference speed of 76.51 FPS. With only 2.8 million parameters, FG-YOLO delivers high performance with optimized computational efficiency, offering a robust solution for real-time fire and smoke detection in safety-critical applications.