Scholarlink

<p>High-throughput, real-time processing is necessary for automated flaw detection of liquid crystal displays (LCDs), but balancing accuracy and latency on devices with constrained resources is still a major challenge. We suggest LCD-YOLO, a computationally efficient framework, to overcome this High-Performance Computing (HPC) limitation. In terms of computational efficiency, the Shared Aware GroupConv Head (SAGHead) combines feature sharing and group operations to eliminate feature redundancy, thereby significantly increasing parallel processing efficiency. The C2f-KW module optimizes the utilization of dynamic parameters, enhancing detection performance and effectively alleviating resource constraints. The Convolution Attention Guided Fusion (CAGF) module enhances multi-scale feature representation while maintaining small-target details to further guarantee detection precision examination. Experiments on the self-constructed LCD defect dataset (LCD-NET) demonstrate that LCD-YOLO reduces floating-point operations (FLOPs) by 45.6% compared to the baseline while maintaining a mean Average Precision (mAP) of 73.7%. Crucially, the model achieves an inference speed of 152 FPS, satisfying the hard real-time constraints of industrial production lines. This validates the proposed method as an effective solution for real-time high-performance industrial inspection.</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

A low-complexity and lightweight detection network for surface defects on liquid crystal display

  • 期刊论文

  • 发表日期：

    2026年2月22日

查看全文
  • Guanqiang Ruan,
  • Zewei Liao,
  • Honggang Shan,
  • Yilin Ni,
  • Long Chen,
  • Xiangdong Kong,
  • Kuo Yang

摘要

High-throughput, real-time processing is necessary for automated flaw detection of liquid crystal displays (LCDs), but balancing accuracy and latency on devices with constrained resources is still a major challenge. We suggest LCD-YOLO, a computationally efficient framework, to overcome this High-Performance Computing (HPC) limitation. In terms of computational efficiency, the Shared Aware GroupConv Head (SAGHead) combines feature sharing and group operations to eliminate feature redundancy, thereby significantly increasing parallel processing efficiency. The C2f-KW module optimizes the utilization of dynamic parameters, enhancing detection performance and effectively alleviating resource constraints. The Convolution Attention Guided Fusion (CAGF) module enhances multi-scale feature representation while maintaining small-target details to further guarantee detection precision examination. Experiments on the self-constructed LCD defect dataset (LCD-NET) demonstrate that LCD-YOLO reduces floating-point operations (FLOPs) by 45.6% compared to the baseline while maintaining a mean Average Precision (mAP) of 73.7%. Crucially, the model achieves an inference speed of 152 FPS, satisfying the hard real-time constraints of industrial production lines. This validates the proposed method as an effective solution for real-time high-performance industrial inspection.