Scholarlink

<p>The semantic segmentation on edge devices is crucial for real-time tasks. However, existing high-precision semantic segmentation models often struggle to be deployed on edge devices with constrained computing power, primarily due to their large parameter counts and high computational resource consumption. This paper proposes a lightweight semantic segmentation model PIDNet-LW which improves PIDNet through model compression and acceleration to make it suitable for edge devices. PIDNet-LW slashes the convolutional stack in the P-branch by 1/3 and redesigns the Basic Block and BottleNeck Block, cutting parameters 47% versus the original PIDNet while still delivering 74.16% mIoU on Aeroscapes, 76.7% mIoU on Cityscapes. Additionally, this paper applies PIDNet-LW to real-time pollutant prediction tasks on edge devices: the model turns vehicles, buildings and vegetation into spatial proxies for traffic emissions, urban-canyon effects and green-space mitigation, directly boosting the accuracy of pollutant-dispersion modeling. Our work provides an effective approach for real-time pollutant prediction tasks on edge devices.</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

PIDNet-LW: lightweight semantic segmentation for edge devices

  • 期刊论文

  • 发表日期：

    2025年11月17日

查看全文
  • Siyuan Qian,
  • Dongfang Zhao,
  • Xun Zhang,
  • Zhentong Gao,
  • Hailong Wang,
  • Zhihao Guo

摘要

The semantic segmentation on edge devices is crucial for real-time tasks. However, existing high-precision semantic segmentation models often struggle to be deployed on edge devices with constrained computing power, primarily due to their large parameter counts and high computational resource consumption. This paper proposes a lightweight semantic segmentation model PIDNet-LW which improves PIDNet through model compression and acceleration to make it suitable for edge devices. PIDNet-LW slashes the convolutional stack in the P-branch by 1/3 and redesigns the Basic Block and BottleNeck Block, cutting parameters 47% versus the original PIDNet while still delivering 74.16% mIoU on Aeroscapes, 76.7% mIoU on Cityscapes. Additionally, this paper applies PIDNet-LW to real-time pollutant prediction tasks on edge devices: the model turns vehicles, buildings and vegetation into spatial proxies for traffic emissions, urban-canyon effects and green-space mitigation, directly boosting the accuracy of pollutant-dispersion modeling. Our work provides an effective approach for real-time pollutant prediction tasks on edge devices.