Scholarlink

<p>Structural cracks pose a significant threat to the safety, serviceability and durability of reinforced concrete structures. Early and accurate crack detection is essential for preventing progressive deterioration and potential structural failure. However, conventional manual inspection methods are time-consuming, labor-intensive and often inconsistent. Recent advances in machine learning (ML) and deep learning (DL) have enabled automated crack detection as a core component of modern structural health monitoring (SHM) systems. Despite the strong performance of DL-based approaches, the influence of image preprocessing on model robustness and generalization remains insufficiently explored. This study presents a systematic evaluation of three commonly used preprocessing techniques: Data Augmentation, Contrast Limited Adaptive Histogram Equalization (CLAHE), and Gaussian blurring, on two DL architectures: ResNet-50 and the Vision Transformer (ViT). A deduplicated version of a publicly available concrete crack dataset was used to reduce redundancy and mitigate evaluation bias. Model performance was evaluated using accuracy, precision, recall and F1-score under consistent training and testing conditions. The results demonstrate that image preprocessing plays a critical role in model reliability and generalization. Data augmentation consistently improved performance across both architectures, while CLAHE achieved the highest overall accuracy, yielding up to a 0.12% improvement for ResNet-50 and emerging as the most effective strategy for this model. In contrast, Gaussian blurring significantly degraded performance for both models, although ViT demonstrated greater resilience under blurred conditions. Unlike prior studies that apply preprocessing heuristically, this work provides a controlled, architecture-aware analysis, offering practical guidance for designing reliable, field-deployable crack detection systems for SHM applications.</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

Enhancing Concrete Crack Detection Using Preprocessing Techniques on a Convolutional Neural Network and a Vision Transformer

  • 期刊论文

  • 发表日期：

    2026年2月24日

查看全文
  • Joshua Ajayi,
  • Oluwafemi Omotayo

摘要

Structural cracks pose a significant threat to the safety, serviceability and durability of reinforced concrete structures. Early and accurate crack detection is essential for preventing progressive deterioration and potential structural failure. However, conventional manual inspection methods are time-consuming, labor-intensive and often inconsistent. Recent advances in machine learning (ML) and deep learning (DL) have enabled automated crack detection as a core component of modern structural health monitoring (SHM) systems. Despite the strong performance of DL-based approaches, the influence of image preprocessing on model robustness and generalization remains insufficiently explored. This study presents a systematic evaluation of three commonly used preprocessing techniques: Data Augmentation, Contrast Limited Adaptive Histogram Equalization (CLAHE), and Gaussian blurring, on two DL architectures: ResNet-50 and the Vision Transformer (ViT). A deduplicated version of a publicly available concrete crack dataset was used to reduce redundancy and mitigate evaluation bias. Model performance was evaluated using accuracy, precision, recall and F1-score under consistent training and testing conditions. The results demonstrate that image preprocessing plays a critical role in model reliability and generalization. Data augmentation consistently improved performance across both architectures, while CLAHE achieved the highest overall accuracy, yielding up to a 0.12% improvement for ResNet-50 and emerging as the most effective strategy for this model. In contrast, Gaussian blurring significantly degraded performance for both models, although ViT demonstrated greater resilience under blurred conditions. Unlike prior studies that apply preprocessing heuristically, this work provides a controlled, architecture-aware analysis, offering practical guidance for designing reliable, field-deployable crack detection systems for SHM applications.