Scholarlink

<p>Surface defects on hot-rolled steel strips are often small, slender, strip-like and low-contrast, so high-frequency details and elongated structures can be lost during conventional convolutional downsampling and feature fusion. To address these issues, we propose Wavelet–Strip-Oriented YOLO11 (WSO-YOLO11), a wavelet–strip prior enhanced network based on YOLO11s. A Wavelet–Space-to-Depth Hybrid Stem (WSH-Stem) is introduced into the shallow backbone to preserve fine-grained textures, and an Orientation-Strip Residual Neck (OSR-Neck) is integrated into the FPN/PAN structure to strengthen the continuous perception of narrow defects. WSO-YOLO11 contains about 9.0M parameters, approximately 4.3% fewer than YOLO11s. On NEU-DET, WSO-YOLO11 improves mAP@0.5 from 73.5% to 75.8%, mAP@0.5:0.95 from 38.2% to 39.7%, and F1 from 68.1% to 69.3%. On GC10-DET, X-SDD and KSDD2, it also improves mAP@0.5, with the largest gain reaching 3.4 percentage points. These results indicate that WSO-YOLO11 achieves moderate improvements in mAP-related metrics while maintaining a compact model size, and shows relatively stable detection performance across multiple steel strip surface defect datasets.</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

WSO-YOLO11: a wavelet–strip-oriented YOLO11 framework for hot-rolled steel strip surface defect detection

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月30日

查看全文
  • Junwu Lin,
  • Linxuan Chen,
  • Cunhan Guo,
  • Xiaofang Wu,
  • Huilin Xu,
  • Na Song

摘要

Surface defects on hot-rolled steel strips are often small, slender, strip-like and low-contrast, so high-frequency details and elongated structures can be lost during conventional convolutional downsampling and feature fusion. To address these issues, we propose Wavelet–Strip-Oriented YOLO11 (WSO-YOLO11), a wavelet–strip prior enhanced network based on YOLO11s. A Wavelet–Space-to-Depth Hybrid Stem (WSH-Stem) is introduced into the shallow backbone to preserve fine-grained textures, and an Orientation-Strip Residual Neck (OSR-Neck) is integrated into the FPN/PAN structure to strengthen the continuous perception of narrow defects. WSO-YOLO11 contains about 9.0M parameters, approximately 4.3% fewer than YOLO11s. On NEU-DET, WSO-YOLO11 improves mAP@0.5 from 73.5% to 75.8%, mAP@0.5:0.95 from 38.2% to 39.7%, and F1 from 68.1% to 69.3%. On GC10-DET, X-SDD and KSDD2, it also improves mAP@0.5, with the largest gain reaching 3.4 percentage points. These results indicate that WSO-YOLO11 achieves moderate improvements in mAP-related metrics while maintaining a compact model size, and shows relatively stable detection performance across multiple steel strip surface defect datasets.