Scholarlink

<p>The deterioration of concrete infrastructure poses significant risks to public safety and economic stability. Traditional manual inspection methods are labor-intensive, subjective, and often limited by access constraints. This study proposes a hybrid framework that integrates a pre-trained Inception-V3 convolutional neural network (CNN) with advanced image processing (IP) techniques for automated crack detection and quantitative characterization in concrete structures. A dataset of 34,000 images, sourced from the SDNET repository and field photography, was used to train and validate the CNN, classifying images into diagonal, horizontal, vertical cracks, and uncracked surfaces. Post classification, the image processing pipeline extracted key geometric parameters including crack angle, width, endpoint length, and actual path length, converting pixel measurements to real world units. Experimental evaluation on site images showed high accuracy, with mean relative errors of 2.35% for angle, 11.23% for width, and 0.80% for endpoint length. The proposed method’s actual path length measurements exceeded endpoint lengths by an average of 14.88%, capturing curvature and irregularity overlooked in straight-line estimates. These results demonstrate the framework’s potential to deliver comprehensive, accurate, and scalable crack assessments, supporting informed maintenance decisions and enhancing the reliability of vision-based structural health monitoring systems.</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

Hybrid CNN and image processing framework for precise characterization of cracks in concrete structures

  • 期刊论文

  • 发表日期：

    2025年9月25日

查看全文
  • Esar Ahmad,
  • Bimalendu Dash,
  • Anurag Tripathi,
  • Padma Mishra,
  • T. C. Manjunath,
  • Krushna Chandra Sethi

摘要

The deterioration of concrete infrastructure poses significant risks to public safety and economic stability. Traditional manual inspection methods are labor-intensive, subjective, and often limited by access constraints. This study proposes a hybrid framework that integrates a pre-trained Inception-V3 convolutional neural network (CNN) with advanced image processing (IP) techniques for automated crack detection and quantitative characterization in concrete structures. A dataset of 34,000 images, sourced from the SDNET repository and field photography, was used to train and validate the CNN, classifying images into diagonal, horizontal, vertical cracks, and uncracked surfaces. Post classification, the image processing pipeline extracted key geometric parameters including crack angle, width, endpoint length, and actual path length, converting pixel measurements to real world units. Experimental evaluation on site images showed high accuracy, with mean relative errors of 2.35% for angle, 11.23% for width, and 0.80% for endpoint length. The proposed method’s actual path length measurements exceeded endpoint lengths by an average of 14.88%, capturing curvature and irregularity overlooked in straight-line estimates. These results demonstrate the framework’s potential to deliver comprehensive, accurate, and scalable crack assessments, supporting informed maintenance decisions and enhancing the reliability of vision-based structural health monitoring systems.