Scholarlink

<p>Wildfires pose significant threats to human safety and the environment, necessitating prompt detection and localization. Deep learning algorithms, particularly those for image segmentation, show promise but often struggle with computational delays, hindering real-time applications. This paper introduces LSNet, a lightweight segmentation network tailored for real-time wildfire detection. LSNet comprises an encoder equipped with a parallel feature multiplier and a residual feature extractor to reduce computational load and expedite segmentation. The decoder incorporates an attentional feature pyramid to enhance segmentation accuracy. Experimental results demonstrate that LSNet achieves a speed of 31 FPS, an AP<sub>50</sub> of 65.1%, and a model size of 274&#xa0;MB during training, outperforming both CNN-based SOLOv2 and transformer-based SegFormer. Testing results further confirm its efficacy with 30 FPS and an AP<sub>50</sub> of 72.3%, meeting the demands of real-time wildfire detection. Thus, LSNet presents a viable solution for timely wildfire identification and localization. Source code is available at: <a href="https://github.com/xiaoyjing/LSNet">https://github.com/xiaoyjing/LSNet</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

Lightweight segmentation network for real-time wildfire detection: LSNet's parallel feature multiplication and attentional fusion

  • 期刊论文

  • 发表日期：

    2025年5月17日

查看全文
  • Cheng Zhang,
  • Zhuoyue Ding,
  • Xiaoying Jing,
  • Lei Huang,
  • Run Ye,
  • Bin Yan,
  • Xiaojia Zhou,
  • Jinhong Guo

摘要

Wildfires pose significant threats to human safety and the environment, necessitating prompt detection and localization. Deep learning algorithms, particularly those for image segmentation, show promise but often struggle with computational delays, hindering real-time applications. This paper introduces LSNet, a lightweight segmentation network tailored for real-time wildfire detection. LSNet comprises an encoder equipped with a parallel feature multiplier and a residual feature extractor to reduce computational load and expedite segmentation. The decoder incorporates an attentional feature pyramid to enhance segmentation accuracy. Experimental results demonstrate that LSNet achieves a speed of 31 FPS, an AP50 of 65.1%, and a model size of 274 MB during training, outperforming both CNN-based SOLOv2 and transformer-based SegFormer. Testing results further confirm its efficacy with 30 FPS and an AP50 of 72.3%, meeting the demands of real-time wildfire detection. Thus, LSNet presents a viable solution for timely wildfire identification and localization. Source code is available at: https://github.com/xiaoyjing/LSNet.