Scholarlink

<p>The automation of road damage detection and classification is vital for enhancing road safety and enabling cost-effective, proactive infrastructure maintenance. This study introduces an advanced transfer learning-based framework designed to achieve precise and efficient road damage classification, validated on the RDD 2022 dataset, which spans six geographically diverse regions—India, Japan, Czech Republic, Norway, the U.S, and China. To ensure both specialization and generalization, the framework employs a two-phase fine-tuning strategy. A region-specific model is trained on country-wise datasets to capture localized damage patterns with high precision, while a generalized model leverages a cross-regional dataset encompassing diverse damage classes from all six countries, enhancing adaptability and ensuring robust performance across varying road conditions. To extract high-dimensional, discriminative representations of road damage images, the framework integrates twelve state-of-the-art deep learning architectures, including VGG16, VGG19, Xception, ResNet50, ResNet101, ResNet152, InceptionV3, DenseNet121, EfficientNetB0, and transformer-based models such as Vision Transformer and Swin Transformer. These extracted features undergo further refinement through K-means clustering, which enhances classification accuracy by structuring the feature space. Additionally, LIME-based feature selection identifies the 100 most salient features, ensuring a balance between computational efficiency and classification performance. The optimized feature vectors serve as inputs for multiple Machine learning classifiers, including SVM, KNN, Decision Trees, Gaussian Naïve Bayes, Logistic Regression, MLP, AdaBoost, XGBoost, Random Forest, LGBM (Light Gradient Boosting Machine), and Extra Trees, ensuring scalable and robust classification across diverse datasets. By integrating deep learning-based feature extraction, clustering-based refinement, and interpretable feature selection, the proposed methodology establishes a new benchmark in automated road maintenance. This comprehensive approach offers a scalable, high-precision solution that effectively balances accuracy, efficiency, and adaptability, making it well-suited for both localized and generalized datasets.</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

An enhanced approach to road damage classification with deep feature extraction and class-specific clustering

  • 期刊论文

  • 发表日期：

    2026年1月10日

查看全文
  • R. Rakshitha,
  • S. Srinath,
  • N. Vinay Kumar,
  • S. Rashmi,
  • B. V. Poornima

摘要

The automation of road damage detection and classification is vital for enhancing road safety and enabling cost-effective, proactive infrastructure maintenance. This study introduces an advanced transfer learning-based framework designed to achieve precise and efficient road damage classification, validated on the RDD 2022 dataset, which spans six geographically diverse regions—India, Japan, Czech Republic, Norway, the U.S, and China. To ensure both specialization and generalization, the framework employs a two-phase fine-tuning strategy. A region-specific model is trained on country-wise datasets to capture localized damage patterns with high precision, while a generalized model leverages a cross-regional dataset encompassing diverse damage classes from all six countries, enhancing adaptability and ensuring robust performance across varying road conditions. To extract high-dimensional, discriminative representations of road damage images, the framework integrates twelve state-of-the-art deep learning architectures, including VGG16, VGG19, Xception, ResNet50, ResNet101, ResNet152, InceptionV3, DenseNet121, EfficientNetB0, and transformer-based models such as Vision Transformer and Swin Transformer. These extracted features undergo further refinement through K-means clustering, which enhances classification accuracy by structuring the feature space. Additionally, LIME-based feature selection identifies the 100 most salient features, ensuring a balance between computational efficiency and classification performance. The optimized feature vectors serve as inputs for multiple Machine learning classifiers, including SVM, KNN, Decision Trees, Gaussian Naïve Bayes, Logistic Regression, MLP, AdaBoost, XGBoost, Random Forest, LGBM (Light Gradient Boosting Machine), and Extra Trees, ensuring scalable and robust classification across diverse datasets. By integrating deep learning-based feature extraction, clustering-based refinement, and interpretable feature selection, the proposed methodology establishes a new benchmark in automated road maintenance. This comprehensive approach offers a scalable, high-precision solution that effectively balances accuracy, efficiency, and adaptability, making it well-suited for both localized and generalized datasets.