Scholarlink

<p>The scarce availability of image datasets for Carbon Fiber Reinforced Polymer (CFRP)-strengthened concrete often results in the underestimation of CFRP defect detection with existing deep learning approaches. To address this challenge, this study introduces a transferable YOLO-DeepCrack model to enable fast and accurate detection of varied types of defects on CFRP-strengthened concrete with small datasets. A novel two-phase transfer learning strategy is presented in this model to inherently convert the learned feature domain from the conventional concrete cracks to the CFRP area defects so that the model can effectively detect CFRP defects even under data-constrained conditions. Compared to the state-of-the-art (SOTA) defect detection models, the proposed method demonstrated outstanding performance across standard evaluation metrics, achieving <i>Precision</i>, <i>Recall</i>, <i>F1-score</i> and <i>mAP@0.5</i> values of 85.9%, 80.9%, 83.3% and 89.9% respectively, while substantially increases the computational efficiency.</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

Developing a transferable YOLO-DeepCrack model for surface defect detection in CFRP-strengthened concrete

  • 期刊论文

  • 发表日期：

    2026年7月20日

查看全文
  • Keyu Chen,
  • Jun Lin,
  • Zilong Wang,
  • Zhexiong Shang

摘要

The scarce availability of image datasets for Carbon Fiber Reinforced Polymer (CFRP)-strengthened concrete often results in the underestimation of CFRP defect detection with existing deep learning approaches. To address this challenge, this study introduces a transferable YOLO-DeepCrack model to enable fast and accurate detection of varied types of defects on CFRP-strengthened concrete with small datasets. A novel two-phase transfer learning strategy is presented in this model to inherently convert the learned feature domain from the conventional concrete cracks to the CFRP area defects so that the model can effectively detect CFRP defects even under data-constrained conditions. Compared to the state-of-the-art (SOTA) defect detection models, the proposed method demonstrated outstanding performance across standard evaluation metrics, achieving Precision, Recall, F1-score and mAP@0.5 values of 85.9%, 80.9%, 83.3% and 89.9% respectively, while substantially increases the computational efficiency.