Scholarlink

<p>To improve accuracy and generalization in early forest fire smoke detection, this paper proposes a novel Wavelet-Attention YOLO (WA-YOLO) scheme based on wavelet convolution and the SimAM attention mechanism. The model integrates wavelet convolution to enhance multi-scale extraction of high- and low-frequency smoke features, and incorporates SimAM to enhance attention on smoke regions for precise localization. A Temporal Directional Spatial Enhancement (TDSE) structure is designed to optimize feature fusion and information flow, while a Dual-Branch Initialization Optimization (DBIO) strategy improves detection head initialization, accelerating convergence and reducing false positives. Experimental results on the both PFE-v8 and try123-v4 datasets show that the proposed WA-YOLO can improve precision by 3.6% and mAP@50% by 1.4% on try123-v4, and achieve 96.5% precision and 86.7% mAP@50% on PFE-v8, demonstrating strong generalization ability and practical value for early forest fire warning. GitHub: <a href="https://github.com/zel4539/WA-YOLO">https://github.com/zel4539/WA-YOLO</a></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

WA-YOLO: a forest fire smoke detection method based on wavelet convolution and SimAM

  • 期刊论文

  • 发表日期：

    2025年11月3日

查看全文
  • Xin Chen,
  • Enliang Zhu,
  • Yaolin Zhu,
  • Yan Fu,
  • Kai Han,
  • Bing Liu

摘要

To improve accuracy and generalization in early forest fire smoke detection, this paper proposes a novel Wavelet-Attention YOLO (WA-YOLO) scheme based on wavelet convolution and the SimAM attention mechanism. The model integrates wavelet convolution to enhance multi-scale extraction of high- and low-frequency smoke features, and incorporates SimAM to enhance attention on smoke regions for precise localization. A Temporal Directional Spatial Enhancement (TDSE) structure is designed to optimize feature fusion and information flow, while a Dual-Branch Initialization Optimization (DBIO) strategy improves detection head initialization, accelerating convergence and reducing false positives. Experimental results on the both PFE-v8 and try123-v4 datasets show that the proposed WA-YOLO can improve precision by 3.6% and mAP@50% by 1.4% on try123-v4, and achieve 96.5% precision and 86.7% mAP@50% on PFE-v8, demonstrating strong generalization ability and practical value for early forest fire warning. GitHub: https://github.com/zel4539/WA-YOLO