Scholarlink

<p>Fiberglass, a critical material in various industrial applications, often exhibits small and inconspicuous defects that are challenging to detect using traditional vision-based methods, especially under few-shot conditions. To address this, we introduce a defect segmentation framework that integrates a Dense Attention Pyramid Network (DAP-Net) with a realistic data synthesis strategy. The synthesis module adaptively fuses limited real samples to generate diverse defect patterns, enriching the training data. DAP-Net employs a dense encoder, improved pyramid decoder, and CBAM attention to enhance cross-scale feature fusion and refine subtle defect localization. Experiments demonstrate state-of-the-art performance with I-F1, P-AP, and P-F1 scores of 94.84%, 73.24%, and 70.93%, respectively, while maintaining lightweight efficiency. This framework offers an effective solution for small-scale defect segmentation in fiberglass inspection. The source code used in this study is available at <a href="https://github.com/SwaggyPinqi12/DAP-Net">https://github.com/SwaggyPinqi12/DAP-Net</a>.</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

Enhancing small-scale fiberglass defect segmentation: a dense attention pyramid network with realistic data augmentation

  • 期刊论文

  • 发表日期：

    2026年2月7日

查看全文
  • Pinqi Cheng,
  • Yanfang Feng,
  • Ziyue Hao,
  • Xuanyin Wang

摘要

Fiberglass, a critical material in various industrial applications, often exhibits small and inconspicuous defects that are challenging to detect using traditional vision-based methods, especially under few-shot conditions. To address this, we introduce a defect segmentation framework that integrates a Dense Attention Pyramid Network (DAP-Net) with a realistic data synthesis strategy. The synthesis module adaptively fuses limited real samples to generate diverse defect patterns, enriching the training data. DAP-Net employs a dense encoder, improved pyramid decoder, and CBAM attention to enhance cross-scale feature fusion and refine subtle defect localization. Experiments demonstrate state-of-the-art performance with I-F1, P-AP, and P-F1 scores of 94.84%, 73.24%, and 70.93%, respectively, while maintaining lightweight efficiency. This framework offers an effective solution for small-scale defect segmentation in fiberglass inspection. The source code used in this study is available at https://github.com/SwaggyPinqi12/DAP-Net.