Scholarlink

<p>Identification of concrete structural cracks at their early stages is essential for safety purposes because these early indicators lead to potential safety risks. A performance analysis of YOLOv8 object detection is provided for automatic concrete crack recognition through different training protocol assessments. The evaluation used actual concrete surface image sets that grew in size from 300 to 600 and finally 1500 samples with crack and non-crack content. YOLOv8n received training using each subset until 10, 30, 50, and 100 epochs for a total of twelve distinct model configurations. The evaluation metrics included Precision, Recall, F1-score, together with mAP@0.5 and mAP@0.5:0.95. Normalized confusion matrices, F1 and PR curves were used with validation predictions for analyzing detection patterns of each model. Model accuracy and generalization reach their peak performance as a result of the combined influence of the training duration and dataset size. A model training with 1500 images for 100 epochs delivered the optimal results, showing Precision 0.9700, Recall 0.9394, and mAP@0.5 of 0.9810 and mAP@0.5:0.95 of 0.8123. The research shows that extending training time past 50 epochs does not provide additional value in performance unless working with small-scale datasets. The research demonstrates that YOLOv8 functions as a viable and efficient tool for automation in crack detection tasks. It lays the groundwork for future enhancements such as severity classification, larger scale dataset integration, and edge deployment for real-time infrastructure monitoring.</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

A performance-driven evaluation of deep learning for concrete crack detection with varying dataset sizes and training epochs: real-world implications for infrastructure monitoring

  • 期刊论文

  • 发表日期：

    2025年5月15日

查看全文
  • Shashi Kumar Bussa,
  • Narendra Kumar Boppana,
  • Bhupesh Deka

摘要

Identification of concrete structural cracks at their early stages is essential for safety purposes because these early indicators lead to potential safety risks. A performance analysis of YOLOv8 object detection is provided for automatic concrete crack recognition through different training protocol assessments. The evaluation used actual concrete surface image sets that grew in size from 300 to 600 and finally 1500 samples with crack and non-crack content. YOLOv8n received training using each subset until 10, 30, 50, and 100 epochs for a total of twelve distinct model configurations. The evaluation metrics included Precision, Recall, F1-score, together with mAP@0.5 and mAP@0.5:0.95. Normalized confusion matrices, F1 and PR curves were used with validation predictions for analyzing detection patterns of each model. Model accuracy and generalization reach their peak performance as a result of the combined influence of the training duration and dataset size. A model training with 1500 images for 100 epochs delivered the optimal results, showing Precision 0.9700, Recall 0.9394, and mAP@0.5 of 0.9810 and mAP@0.5:0.95 of 0.8123. The research shows that extending training time past 50 epochs does not provide additional value in performance unless working with small-scale datasets. The research demonstrates that YOLOv8 functions as a viable and efficient tool for automation in crack detection tasks. It lays the groundwork for future enhancements such as severity classification, larger scale dataset integration, and edge deployment for real-time infrastructure monitoring.