Scholarlink

<p>Rapid and accurate detection of photovoltaic (PV) electroluminescence (EL) defects is essential for quality control in the production process. However, the high computational demands of GPU-based models limit the deployment on edge devices. To achieve efficient deployment on resource-constrained embedded platforms, A lightweight detection framework Tiny-YOLOv8 is proposed. It employs StarNet as the backbone, C3x in neck, and designs a Lightweight Shared Convolutional Detection Head (LSCD) to reduce model complexity. During the inference process, Layer-adaptive Magnitude Pruning (LAMP) and Knowledge Distillation (KD) are combined to compensate for the accuracy loss caused by the lightweight design. In the deployment stage, the output processing workflow is optimized to solve the incompatibility between RK3588 platform and DFL algorithm. Meanwhile, a producer–consumer parallel inference mode is modified to improve the utilization of the three-core NPU. Experimental results demonstrate that Tiny-YOLOv8 reduces Params by 86.1%, GFLOPs by 78.0%, and increases FPS by 124.6%, with only 0.9% decrease of mAP@0.5–0.95 on GPU. On RK3588 platform, FPS improves by 51% on 6 threads. The proposed approach not only mitigates dependence of deep learning algorithms on GPU, but also provides an effective solution for real-time PV defect inspection in industrial 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

Tiny-YOLOv8: an embedded deployment approach engineering realization for electroluminescence image detection

  • 期刊论文

  • 发表日期：

    2026年2月6日

查看全文
  • Sheng Ding,
  • Ran Song,
  • Congyan Chen

摘要

Rapid and accurate detection of photovoltaic (PV) electroluminescence (EL) defects is essential for quality control in the production process. However, the high computational demands of GPU-based models limit the deployment on edge devices. To achieve efficient deployment on resource-constrained embedded platforms, A lightweight detection framework Tiny-YOLOv8 is proposed. It employs StarNet as the backbone, C3x in neck, and designs a Lightweight Shared Convolutional Detection Head (LSCD) to reduce model complexity. During the inference process, Layer-adaptive Magnitude Pruning (LAMP) and Knowledge Distillation (KD) are combined to compensate for the accuracy loss caused by the lightweight design. In the deployment stage, the output processing workflow is optimized to solve the incompatibility between RK3588 platform and DFL algorithm. Meanwhile, a producer–consumer parallel inference mode is modified to improve the utilization of the three-core NPU. Experimental results demonstrate that Tiny-YOLOv8 reduces Params by 86.1%, GFLOPs by 78.0%, and increases FPS by 124.6%, with only 0.9% decrease of mAP@0.5–0.95 on GPU. On RK3588 platform, FPS improves by 51% on 6 threads. The proposed approach not only mitigates dependence of deep learning algorithms on GPU, but also provides an effective solution for real-time PV defect inspection in industrial edge devices.