Scholarlink

<p>Steel surface defects inevitably occur during production due to process and equipment factors. Steel manufacturers collect and annotate samples of these defects to construct datasets for training inspection models used in production. However, novel defect types can emerge during ongoing operations. To address this challenge, we propose FS-SSDD. FS-SSDD enables the training of viable detection models using very limited samples of new defect types. Our core innovation lies in a hybrid module effectively fusing CNN and Transformer architectures. This module integrates with the backbone and feature fusion networks, forming a Hybrid Backbone and a Hybrid Feature Fusion Network. These networks significantly enhance the model’s capability to model intricate features of steel surface defects. Furthermore, to mitigate overfitting, a prevalent issue in few-shot training, we introduce an Extended Region Proposal Network (ERPN). We also propose the Adaptive Contrastive Proposal Encoding loss to reduce the adverse impact of noisy samples on model training. For optimizing the detection head, FS-SSDD employs Focal Loss for classification and L1 Loss for bounding box regression. The candidate proposals generated by the ERPN are refined using Soft Non-Maximum Suppression. Comprehensive experiments conducted on the NEU-DET and GC10-DET datasets demonstrate that FS-SSDD possesses remarkable performance. The implementation code is available at: <a href="https://github.com/ghlerrix/FS-SSDD">https://github.com/ghlerrix/FS-SSDD</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

FS-SSDD: A Few-shot Steel Surface Defect Detection Network

  • 期刊论文

  • 发表日期：

    2025年10月12日

查看全文
  • Guanghu Liu,
  • Maoxiang Chu,
  • Rongfen Gong,
  • Peinan Zong

摘要

Steel surface defects inevitably occur during production due to process and equipment factors. Steel manufacturers collect and annotate samples of these defects to construct datasets for training inspection models used in production. However, novel defect types can emerge during ongoing operations. To address this challenge, we propose FS-SSDD. FS-SSDD enables the training of viable detection models using very limited samples of new defect types. Our core innovation lies in a hybrid module effectively fusing CNN and Transformer architectures. This module integrates with the backbone and feature fusion networks, forming a Hybrid Backbone and a Hybrid Feature Fusion Network. These networks significantly enhance the model’s capability to model intricate features of steel surface defects. Furthermore, to mitigate overfitting, a prevalent issue in few-shot training, we introduce an Extended Region Proposal Network (ERPN). We also propose the Adaptive Contrastive Proposal Encoding loss to reduce the adverse impact of noisy samples on model training. For optimizing the detection head, FS-SSDD employs Focal Loss for classification and L1 Loss for bounding box regression. The candidate proposals generated by the ERPN are refined using Soft Non-Maximum Suppression. Comprehensive experiments conducted on the NEU-DET and GC10-DET datasets demonstrate that FS-SSDD possesses remarkable performance. The implementation code is available at: https://github.com/ghlerrix/FS-SSDD.