Evaluation of deep learning models for classification of asphalt pavement distresses

Transfer learning (TL) offers a convenient methodology for exploiting the capability of deep convolutional neural networks (DCNNs) for many image classification tasks including the classification of pavement distresses. Seven state-of-the-art DCNNs were retrained to classify asphalt pavement distresses grouped into eight classes using TL techniques. The aim was to evaluate the predictive performances of the selected DCNNs in order to provide some guidelines on selection of DCNNs for pavement application. The results show some existing DCNN's are better than others for developing pavement distress classification models using the specific TL approach adopted in the study. The predictive ability of each model varied depending on distress class as some models with very low overall accuracy showed excellent results for individual distress class(s). Based on a combination of various performance metrics including F1-score, area under ROC curve, optimal operating threshold, training time, and model size, the best performing network had a relative score that was found to be significantly higher than the next two top-performing models. The best-performing networks were characterised by lower proportions of false negative values, low ambiguity scores, and well-defined t-SNE clusters that showed clear separation between the eight different pavement distress classes considered.

Automated summary

This study explores the application of transfer learning in retraining seven state-of-the-art deep convolutional neural networks (DCNNs) to classify asphalt pavement distresses. The results reveal that certain DCNNs outperform others when adopting the specific transfer learning approach. Based on multiple performance metrics, the best-performing network exhibits lower false negative rates, lower ambiguity scores, and well-defined t-SNE clusters.