Scholarlink

<p>Accurate detection of surface defects in industrial steel production is crucial for ensuring product quality and sustaining manufacturing efficiency. However, existing deep learning-based detection models often struggle with imprecise localization, frequent misclassifications, and ambiguity between defect regions and complex background patterns, which degrades detection performance in challenging defect scenarios. To address these challenges, this study introduces an enhanced model for steel surface defect detection that integrates multi-scale feature fusion with a poly-kernel inception network. The model incorporates an attention-based intra-scale feature extraction (AIFI) module to enhance feature correlation, a poly-kernel inception (PKI) block for multi-scale feature extraction, and a generalized feature pyramid network (GFPN) structure for efficient feature integration. Experimental evaluations on NEU-DET and GC10-DET datasets demonstrate the superior performance of the proposed model. On NEU-DET, the mean average precision reaches 82.4%, representing a 5.7% improvement over the baseline YOLOv11s model. On GC10-DET, the mean average precision achieves 67.6%, with an improvement of 4.4%. In addition, the probability of simultaneous localization and classification errors is significantly reduced, decreasing from 0.04 to 0.01 on NEU-DET and from 1.52 to 0.86 on GC10-DET. These results confirm the enhanced detection accuracy and robustness of the proposed model, particularly in handling challenging samples such as low-contrast and blurred images. Code is available at <a href="https://github.com/liaojiayi-dot/APG-YOLO">https://github.com/liaojiayi-dot/APG-YOLO</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

Enhanced multi-scale feature fusion for accurate steel surface defect detection

  • 期刊论文

  • 发表日期：

    2026年7月11日

查看全文
  • Jiayi Liao,
  • Deguang Wang,
  • Jianfang Liang,
  • Ming Yang,
  • Chengbin Liang

摘要

Accurate detection of surface defects in industrial steel production is crucial for ensuring product quality and sustaining manufacturing efficiency. However, existing deep learning-based detection models often struggle with imprecise localization, frequent misclassifications, and ambiguity between defect regions and complex background patterns, which degrades detection performance in challenging defect scenarios. To address these challenges, this study introduces an enhanced model for steel surface defect detection that integrates multi-scale feature fusion with a poly-kernel inception network. The model incorporates an attention-based intra-scale feature extraction (AIFI) module to enhance feature correlation, a poly-kernel inception (PKI) block for multi-scale feature extraction, and a generalized feature pyramid network (GFPN) structure for efficient feature integration. Experimental evaluations on NEU-DET and GC10-DET datasets demonstrate the superior performance of the proposed model. On NEU-DET, the mean average precision reaches 82.4%, representing a 5.7% improvement over the baseline YOLOv11s model. On GC10-DET, the mean average precision achieves 67.6%, with an improvement of 4.4%. In addition, the probability of simultaneous localization and classification errors is significantly reduced, decreasing from 0.04 to 0.01 on NEU-DET and from 1.52 to 0.86 on GC10-DET. These results confirm the enhanced detection accuracy and robustness of the proposed model, particularly in handling challenging samples such as low-contrast and blurred images. Code is available at https://github.com/liaojiayi-dot/APG-YOLO.