Scholarlink

<p>Identifying surface cracks (SC) is essential to guarantee the safety and longevity of infrastructure, including bridges, roads, and buildings. Traditional methods are error-prone, time-consuming, and expensive. An automated SC detection and classification system using an advanced deep learning (DL)-based image processing technique can address the existing limitations in classifying SC. In this study, the authors present a DL-based model to detect and classify concrete and asphalt SC. An advanced hybrid model integrating MobileNet V3, LeViT, and an enhanced Linformer was employed for feature extraction. A gated feature fusion mechanism was used to fuse the crucial features to improve the classification performance. A fine-tuned Kolmogorov-Arnold Networks (KANs) was employed to classify the features into negative and positive classes. To enable the model’s interpretability, the authors integrated Shapley Additive Explanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) values with the proposed model. The experimental outcomes demonstrate the model’s generalization performance, achieving a remarkable accuracy of 99.5% with a low computational loss of 0.09 on the concrete crack conglomerate dataset, outperforming existing models. The findings underscore the model’s generalization ability across diverse datasets. By deploying the proposed model on mobile and edge devices, continuous and autonomous infrastructure monitoring can be performed. Using the proposed model, the reliance on manual inspections can be minimized. The proposed model offers a potential solution for advancing smart infrastructure maintenance 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

Vision transformers- Kolmogorov–Arnold networks-based consumer driven surface cracks classification model

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年2月15日

查看全文
  • Abdul Rahaman Wahab Sait,
  • Suresh Sankaranarayanan,
  • Yang Yu

摘要

Identifying surface cracks (SC) is essential to guarantee the safety and longevity of infrastructure, including bridges, roads, and buildings. Traditional methods are error-prone, time-consuming, and expensive. An automated SC detection and classification system using an advanced deep learning (DL)-based image processing technique can address the existing limitations in classifying SC. In this study, the authors present a DL-based model to detect and classify concrete and asphalt SC. An advanced hybrid model integrating MobileNet V3, LeViT, and an enhanced Linformer was employed for feature extraction. A gated feature fusion mechanism was used to fuse the crucial features to improve the classification performance. A fine-tuned Kolmogorov-Arnold Networks (KANs) was employed to classify the features into negative and positive classes. To enable the model’s interpretability, the authors integrated Shapley Additive Explanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) values with the proposed model. The experimental outcomes demonstrate the model’s generalization performance, achieving a remarkable accuracy of 99.5% with a low computational loss of 0.09 on the concrete crack conglomerate dataset, outperforming existing models. The findings underscore the model’s generalization ability across diverse datasets. By deploying the proposed model on mobile and edge devices, continuous and autonomous infrastructure monitoring can be performed. Using the proposed model, the reliance on manual inspections can be minimized. The proposed model offers a potential solution for advancing smart infrastructure maintenance systems.