Scholarlink

<p>Transmission line insulators are critical components for stable grid operation, but their condition detection is often compromised by motion-induced image blurring during UAV inspections, leading to reduced defect analysis accuracy. To address this, this paper proposes a lightweight joint optimization framework for image restoration and target detection. In image restoration, the generator’s backbone network in DeblurGANv2 is replaced with a Ghost module to reduce model parameters. An innovative Constrained Principal Component Analysis Attention (CPCA-Attention) mechanism is introduced, which dynamically adjusts channel weights to suppress redundant features and spatial weights to enhance edge details in feature maps, resulting in a PSNR of 31.5 dB on the dataset. For target detection, a lightweight YOLO-Lite model is developed, incorporating a C2f-Faster-EMA module to capture long-range pixel dependencies while reducing parameters and floating-point operations, thereby improving the global perception of small targets. Additionally, a Shape-IoU loss function is employed to accelerate convergence by prioritizing high-quality anchors, and a SEAM network is integrated into the detection head to better handle occlusions. The improved model demonstrates enhanced robustness in detecting small targets, occluded targets, and targets in complex backgrounds, achieving a mean Average Precision (mAP) of 97.5%, with FLOPS reduced by 19.7% and parameters decreased by 16.7%.</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

YOLO-Lite: Lightweight Object Detection Algorithm Design and Optimization in Power System Inspection

  • 期刊论文

  • 发表日期：

    2025年6月9日

查看全文
  • Lihui Lu,
  • Li Liu,
  • Jiaqi Chen,
  • Jianguo Liu,
  • Xinyuan Wu

摘要

Transmission line insulators are critical components for stable grid operation, but their condition detection is often compromised by motion-induced image blurring during UAV inspections, leading to reduced defect analysis accuracy. To address this, this paper proposes a lightweight joint optimization framework for image restoration and target detection. In image restoration, the generator’s backbone network in DeblurGANv2 is replaced with a Ghost module to reduce model parameters. An innovative Constrained Principal Component Analysis Attention (CPCA-Attention) mechanism is introduced, which dynamically adjusts channel weights to suppress redundant features and spatial weights to enhance edge details in feature maps, resulting in a PSNR of 31.5 dB on the dataset. For target detection, a lightweight YOLO-Lite model is developed, incorporating a C2f-Faster-EMA module to capture long-range pixel dependencies while reducing parameters and floating-point operations, thereby improving the global perception of small targets. Additionally, a Shape-IoU loss function is employed to accelerate convergence by prioritizing high-quality anchors, and a SEAM network is integrated into the detection head to better handle occlusions. The improved model demonstrates enhanced robustness in detecting small targets, occluded targets, and targets in complex backgrounds, achieving a mean Average Precision (mAP) of 97.5%, with FLOPS reduced by 19.7% and parameters decreased by 16.7%.