Scholarlink

<p>Concrete crack detection and segmentation play an important role in the structural health assessment of civil infrastructure. This study investigates the performance of YOLOv11 and YOLOv26 models for concrete crack detection and segmentation under different input image resolutions. Experiments were conducted using four model variants, namely nano (n), small (s), medium (m), and large (l), each trained for 30 epochs. Model performance was evaluated using mAP@0.5, mAP@0.5:0.95, Dice score, mIoU, as well as training and inference time. The results show that YOLOv11 consistently achieves higher detection performance across most configurations, particularly at a resolution of 416 <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\times \)</EquationSource> <EquationSource Format="MATHML"><math> <mo>×</mo> </math></EquationSource> </InlineEquation> 416. In contrast, YOLOv26 demonstrates more stable segmentation performance, maintaining consistent Dice and mIoU values across different resolutions. Increasing the input resolution beyond 416 <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\times \)</EquationSource> <EquationSource Format="MATHML"><math> <mo>×</mo> </math></EquationSource> </InlineEquation> 416 does not lead to significant performance improvements and instead introduces higher computational cost. The findings indicate that a resolution of 416 <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(\times \)</EquationSource> <EquationSource Format="MATHML"><math> <mo>×</mo> </math></EquationSource> </InlineEquation> 416 provides the best balance between accuracy and efficiency for both models. YOLOv11 is more suitable for detection-focused and real-time applications, while YOLOv26 offers a more robust alternative for segmentation tasks requiring stable performance.</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

Exploring the effect of resizing images in concrete crack detection and segmentation

  • 期刊论文

  • 发表日期：

    2026年6月25日

查看全文
  • Mahmud Isnan,
  • Dede Fauzi,
  • Hakas Prayuda

摘要

Concrete crack detection and segmentation play an important role in the structural health assessment of civil infrastructure. This study investigates the performance of YOLOv11 and YOLOv26 models for concrete crack detection and segmentation under different input image resolutions. Experiments were conducted using four model variants, namely nano (n), small (s), medium (m), and large (l), each trained for 30 epochs. Model performance was evaluated using mAP@0.5, mAP@0.5:0.95, Dice score, mIoU, as well as training and inference time. The results show that YOLOv11 consistently achieves higher detection performance across most configurations, particularly at a resolution of 416 \(\times \) × 416. In contrast, YOLOv26 demonstrates more stable segmentation performance, maintaining consistent Dice and mIoU values across different resolutions. Increasing the input resolution beyond 416 \(\times \) × 416 does not lead to significant performance improvements and instead introduces higher computational cost. The findings indicate that a resolution of 416 \(\times \) × 416 provides the best balance between accuracy and efficiency for both models. YOLOv11 is more suitable for detection-focused and real-time applications, while YOLOv26 offers a more robust alternative for segmentation tasks requiring stable performance.