Scholarlink

<p>Accurate discrimination between visually similar deterioration patterns is essential for reliable decision-making in building pathology and rehabilitation planning. Conventional deep-learning studies emphasise model architecture but often overlook the domain-specific visual characteristics that drive misclassification in real inspections. This study introduces a civil-engineering–guided training framework comprising (i) non-destructive quality assessment, (ii) a Civil-Engineering–Aware Augmentation pipeline (CEAA), (iii) diagnostic-preserving stratified partitioning (CEASS), and (iv) a two-stage progressive learning strategy. The approach is evaluated using the BD3 dataset containing seven deterioration types from reinforced-concrete structures. CEAA enriches defect variability and produces a balanced 14,000-image dataset, while CEASS maintains diagnostic sensitivity across training, validation and testing splits. Progressive learning enables each model to first acquire authentic deterioration morphology and subsequently refine fine-grained defect boundaries.Three architectures AlexNet, MobileNetV2 and ViT-Patch16 were trained using architecture-appropriate configurations and evaluated on 2100 held-out validation images. The framework substantially reduced misclassification without requiring architectural modification, achieving final accuracies of 99.24%, 99.76% and 99.86% respectively, with only 16, 5 and 4 misclassifications. These improvements arise from enhanced boundary representation, increased intra-class variability and stabilised convergence, collectively producing a generalisable performance gain. The proposed strategy offers a practical, reproducible pipeline for high-accuracy defect identification and supports broader applications in structural pathology and rehabilitation workflows.</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 civil engineering–aware progressive learning framework for reliable diagnosis of concrete surface pathologies

  • 期刊论文

  • 发表日期：

    2026年2月6日

查看全文
  • S. Parvathy,
  • S. Sivasankar

摘要

Accurate discrimination between visually similar deterioration patterns is essential for reliable decision-making in building pathology and rehabilitation planning. Conventional deep-learning studies emphasise model architecture but often overlook the domain-specific visual characteristics that drive misclassification in real inspections. This study introduces a civil-engineering–guided training framework comprising (i) non-destructive quality assessment, (ii) a Civil-Engineering–Aware Augmentation pipeline (CEAA), (iii) diagnostic-preserving stratified partitioning (CEASS), and (iv) a two-stage progressive learning strategy. The approach is evaluated using the BD3 dataset containing seven deterioration types from reinforced-concrete structures. CEAA enriches defect variability and produces a balanced 14,000-image dataset, while CEASS maintains diagnostic sensitivity across training, validation and testing splits. Progressive learning enables each model to first acquire authentic deterioration morphology and subsequently refine fine-grained defect boundaries.Three architectures AlexNet, MobileNetV2 and ViT-Patch16 were trained using architecture-appropriate configurations and evaluated on 2100 held-out validation images. The framework substantially reduced misclassification without requiring architectural modification, achieving final accuracies of 99.24%, 99.76% and 99.86% respectively, with only 16, 5 and 4 misclassifications. These improvements arise from enhanced boundary representation, increased intra-class variability and stabilised convergence, collectively producing a generalisable performance gain. The proposed strategy offers a practical, reproducible pipeline for high-accuracy defect identification and supports broader applications in structural pathology and rehabilitation workflows.