Scholarlink

<p>Real-time detection of surface defects in industrial products is pivotal for the efficiency of industrial production processes. Over the last decade, deep learning-based object detection algorithms have shown their remarkable performance in this field, especially those recently suggested transformer-based methods. However, they often face challenges related to high computational complexity and substantial memory usage. To address these issues, we propose a Faster Real-Time Detection Transformer (FRT-DETR), designed explicitly for detecting surface defects in industrial products. Our findings highlight that the high computational load of RT-DETR originates from convolutional layers from multiple RepConv blocks within the RepC3 module. To address this issue, we integrate an aggregated SaE layer with the original RepC3 module, aiming to develop a more streamlined SaERepC3 module. Moreover, we introduce the IoU-aware query mechanism to select higher quality queries from boxes generated by the encoder. Additionally, we incorporate a newly designed Multi-Scale Detail Integration (MSDI) module, which employs skip connections to enhance the integration of details across scales, thus enhancing the detection accuracy of the model. From experimental results on the NEU-DET dataset and DeepPCB dataset, our model achieves an impressive inference speed, making it highly effective for real-time surface defect detection tasks.</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

FRT-DETR: faster real-time end-to-end detector for industrial surface defects based on transformer

  • 期刊论文

  • 发表日期：

    2025年9月3日

查看全文
  • Shuang Xiang,
  • Chong Fu,
  • Wei Song,
  • Xingwei Wang,
  • Junxin Chen,
  • Chiu-Wing Sham

摘要

Real-time detection of surface defects in industrial products is pivotal for the efficiency of industrial production processes. Over the last decade, deep learning-based object detection algorithms have shown their remarkable performance in this field, especially those recently suggested transformer-based methods. However, they often face challenges related to high computational complexity and substantial memory usage. To address these issues, we propose a Faster Real-Time Detection Transformer (FRT-DETR), designed explicitly for detecting surface defects in industrial products. Our findings highlight that the high computational load of RT-DETR originates from convolutional layers from multiple RepConv blocks within the RepC3 module. To address this issue, we integrate an aggregated SaE layer with the original RepC3 module, aiming to develop a more streamlined SaERepC3 module. Moreover, we introduce the IoU-aware query mechanism to select higher quality queries from boxes generated by the encoder. Additionally, we incorporate a newly designed Multi-Scale Detail Integration (MSDI) module, which employs skip connections to enhance the integration of details across scales, thus enhancing the detection accuracy of the model. From experimental results on the NEU-DET dataset and DeepPCB dataset, our model achieves an impressive inference speed, making it highly effective for real-time surface defect detection tasks.