Scholarlink

<p>Indoor small-target fires exhibit minimal flame pixels and weak visual features, making their reliable detection a core challenge for early fire warning systems. This study aims to develop a detection model that achieves an optimal balance between high sensitivity and low computational cost for this critical task. Based on YOLOv8n, we first integrated a self-designed Parallel Convolutional Block Attention Module (ParallelCBAM) into the SPPF structure to enhance tiny flame feature extraction and suppress background interference. Secondly, we devised a Difference-Map-Guided feature enhancement module that leverages image self-reconstruction to amplify the saliency of flame regions. On our self-built dataset, the improved model increased the recall and mAP50 by 2.5% and 1.3%, respectively, over the original YOLOv8n. Evaluations on benchmark datasets showed our model significantly reduces parameters while maintaining competitive accuracy. Compared to other literature models, it holds less parameters with slightly improved precision and mAP50. The results validate the effectiveness and practicality of the proposed method in enhancing the reliability of small-target fire perception and facilitating deployment in real-world 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

RDP-YOLO: Reconstruction Difference-Map-Guided YOLO for Indoor Small Flames

  • 期刊论文

  • 发表日期：

    2026年6月27日

查看全文
  • Chunfeng Song,
  • Yan Li,
  • Xueyi Liu

摘要

Indoor small-target fires exhibit minimal flame pixels and weak visual features, making their reliable detection a core challenge for early fire warning systems. This study aims to develop a detection model that achieves an optimal balance between high sensitivity and low computational cost for this critical task. Based on YOLOv8n, we first integrated a self-designed Parallel Convolutional Block Attention Module (ParallelCBAM) into the SPPF structure to enhance tiny flame feature extraction and suppress background interference. Secondly, we devised a Difference-Map-Guided feature enhancement module that leverages image self-reconstruction to amplify the saliency of flame regions. On our self-built dataset, the improved model increased the recall and mAP50 by 2.5% and 1.3%, respectively, over the original YOLOv8n. Evaluations on benchmark datasets showed our model significantly reduces parameters while maintaining competitive accuracy. Compared to other literature models, it holds less parameters with slightly improved precision and mAP50. The results validate the effectiveness and practicality of the proposed method in enhancing the reliability of small-target fire perception and facilitating deployment in real-world applications.