Scholarlink

<p>In high-precision manufacturing, detecting small and complex defects is crucial for ensuring product quality and production efficiency. Industrial micro-defect detection is especially important in the mass production of micro-parts. Traditional methods struggle with accuracy and speed due to the small size, complexity, and high density of the targets. To address these challenges, this paper proposes a transformer-based method using glass beads as a case study, named Glass Bead Detection with Transformers (GLB-DETR). First, a partially convolutional feature extraction network is designed to improve detection accuracy for small targets while reducing model complexity. Next, an optimized hybrid encoder module is introduced to enhance cross-scale feature fusion and real-time detection efficiency. Finally, an improved shape intersection-over-union loss function is applied to improve localization accuracy for small targets. Experimental results show that GLB-DETR outperforms the Real-Time Detection Transformer and You Only Look Once models in both accuracy and real-time performance. Specifically, the mean Average Precision has improved by at least 3.06%, the number of model parameters has been reduced by at least 18.1%, and the detection speed has increased by at least 1.06%. The method also demonstrates strong generalization ability on public datasets, further validating its effectiveness and superiority.</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

An improved transformer-based approach for industrial micro-object defect detection

  • 期刊论文

  • 发表日期：

    2025年9月10日

查看全文
  • Shan Gao,
  • YuMeng Wang,
  • ZongFang Ma

摘要

In high-precision manufacturing, detecting small and complex defects is crucial for ensuring product quality and production efficiency. Industrial micro-defect detection is especially important in the mass production of micro-parts. Traditional methods struggle with accuracy and speed due to the small size, complexity, and high density of the targets. To address these challenges, this paper proposes a transformer-based method using glass beads as a case study, named Glass Bead Detection with Transformers (GLB-DETR). First, a partially convolutional feature extraction network is designed to improve detection accuracy for small targets while reducing model complexity. Next, an optimized hybrid encoder module is introduced to enhance cross-scale feature fusion and real-time detection efficiency. Finally, an improved shape intersection-over-union loss function is applied to improve localization accuracy for small targets. Experimental results show that GLB-DETR outperforms the Real-Time Detection Transformer and You Only Look Once models in both accuracy and real-time performance. Specifically, the mean Average Precision has improved by at least 3.06%, the number of model parameters has been reduced by at least 18.1%, and the detection speed has increased by at least 1.06%. The method also demonstrates strong generalization ability on public datasets, further validating its effectiveness and superiority.