Scholarlink

<p>Steel is an essential material in industrial production. However, high temperatures, intense light, and heavy smoke and dust at production sites often result in blurred images and complex background information in steel surface defect images. To address this issue, this paper proposes YOLO-CESL, a method specifically designed for steel surface defect detection. The proposed network reconstructs spatial features through multilevel feature fusion and enhances the model’s contextual awareness, enabling it to better focus on target objects. Additionally, considering that blurred images contain fewer detailed features and pose difficulties in feature extraction, this paper proposes replacing the pooling-based downsampling strategy with a feature map slicing and recombination approach. It employs multiscale feature extraction to enrich feature representation. Finally, by integrating the Large Separable Kernel Attention, the model is guided to pay more attention to the shape of the target objects. Experimental results show that YOLO-CESL achieves mAP@50 values of 83.1%, 73.9% and 55.3% on the NEU-DET, GC10-DET and SSDD public datasets, respectively, representing an improvement of 5.2%, 6.3% and 4.9% over the baseline YOLOv8. This clear margin of improvement demonstrates the effectiveness and superior detection capability of the proposed model for steel surface defect 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

Efficient Detection Algorithm of Steel Surface Defects Based on YOLOv8

  • 期刊论文

  • 发表日期：

    2026年3月16日

查看全文
  • Shiqi Liu,
  • Xincheng Miao,
  • Ye Zhou

摘要

Steel is an essential material in industrial production. However, high temperatures, intense light, and heavy smoke and dust at production sites often result in blurred images and complex background information in steel surface defect images. To address this issue, this paper proposes YOLO-CESL, a method specifically designed for steel surface defect detection. The proposed network reconstructs spatial features through multilevel feature fusion and enhances the model’s contextual awareness, enabling it to better focus on target objects. Additionally, considering that blurred images contain fewer detailed features and pose difficulties in feature extraction, this paper proposes replacing the pooling-based downsampling strategy with a feature map slicing and recombination approach. It employs multiscale feature extraction to enrich feature representation. Finally, by integrating the Large Separable Kernel Attention, the model is guided to pay more attention to the shape of the target objects. Experimental results show that YOLO-CESL achieves mAP@50 values of 83.1%, 73.9% and 55.3% on the NEU-DET, GC10-DET and SSDD public datasets, respectively, representing an improvement of 5.2%, 6.3% and 4.9% over the baseline YOLOv8. This clear margin of improvement demonstrates the effectiveness and superior detection capability of the proposed model for steel surface defect tasks.