Scholarlink

<p>Fine-grained image signal representation remains challenging when defects exhibit small scale and high inter-class similarity. This problem is particularly evident in industrial surface defect images, where subtle texture variations must be accurately captured for reliable classification. To address this issue, this paper proposes a high-precision feature extraction and classification network named R-CIRK. The network introduces two key improvements in feature extraction and classification mechanisms. Firstly, a stable and robust feature extraction unit (SRFE) and a CBAM-Inverted Residual block (CIR) focusing on fine-grained regions are designed to enhance feature representation and discrimination capability. Secondly, a knowledge-augmented classifier based on the Kolmogorov–Arnold Network (KAN-classifier) is developed. This component improves model generalization and classification performance through nonlinear modeling. Experiments conducted on a self-constructed foil defect dataset demonstrate that R-CIRK achieves an accuracy of 0.979, outperforming the baseline model. These results demonstrate the effectiveness of the proposed method for fine-grained image signal classification and nonlinear feature representation.</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

R-CIRK: A High-Precision Feature Extraction and Classification Network for Industrial Surface Defects

  • 期刊论文

  • 发表日期：

    2026年5月29日

查看全文
  • Linyuan Shi,
  • Yaoyu Shen,
  • Liyuan Lin,
  • Leguang Wang,
  • Xiaoyu Li,
  • Guanhua Qiao,
  • Chaofan Wang,
  • Yonggang Yang,
  • Weibin Zhou

摘要

Fine-grained image signal representation remains challenging when defects exhibit small scale and high inter-class similarity. This problem is particularly evident in industrial surface defect images, where subtle texture variations must be accurately captured for reliable classification. To address this issue, this paper proposes a high-precision feature extraction and classification network named R-CIRK. The network introduces two key improvements in feature extraction and classification mechanisms. Firstly, a stable and robust feature extraction unit (SRFE) and a CBAM-Inverted Residual block (CIR) focusing on fine-grained regions are designed to enhance feature representation and discrimination capability. Secondly, a knowledge-augmented classifier based on the Kolmogorov–Arnold Network (KAN-classifier) is developed. This component improves model generalization and classification performance through nonlinear modeling. Experiments conducted on a self-constructed foil defect dataset demonstrate that R-CIRK achieves an accuracy of 0.979, outperforming the baseline model. These results demonstrate the effectiveness of the proposed method for fine-grained image signal classification and nonlinear feature representation.