Automatic classification of failures of rigid pavements using artificial intelligence techniques

Abstract

We present a novel approach for automatically classifying rigid pavement surface conditions using artificial neural networks (ANNs) and image processing techniques. From a dataset of 5,046 images captured with a mobile mapping system developed by the Universidad Distrital Francisco José de Caldas, 798 images were manually selected and categorized according to the Pavement Condition Index (PCI). These images served to train and to validate four ANN architectures: VGG16, ResNet50, EfficientNetB0, and a custom model referred to as UDM. The models were assessed in terms of accuracy, loss, training time, precision, and recall. The results indicate that ResNet50 and EfficientNetB0 achieved the highest performance, surpassing human evaluators. In particular, EfficientNetB0 demonstrated strong performance with lower computational demands. A comparison with two groups of human evaluators (junior and senior) showed that the ANNs provided greater consistency and accuracy, whereas the human evaluators tended to underestimate pavement conditions and exhibited higher response variability. These findings underscore the potential of artificial intelligence to enhance and standardize road inspection processes by minimizing the subjectivity inherent in manual assessments.