Distress detection in road pavements using neural networks

Abstract

Combining Computer Vision (CV) and Anomaly Detection (AD), there is a convergence of methodologies using convolutional layers in AD architectures, which we consider an innovation in the field. The main goal of this work is to present different Artificial Neural Networks (ANN) architectures, applying them to distress detection in road pavements and comparing the results obtained in each approach. The experimented methods for AD in images include a binary classifier as a baseline, an Autoencoder (AE) and a Variational Autoencoder (VAE). Supervised and unsupervised practises are also compared, proving their utility in scenarios where there is no labelled data available. Using the VAE model in a supervised setting, it presents an excellent distinction between good and bad pavement. When labelled data is not available, using the AE model and the distribution of similarities of good pavement reconstructions to calculate the threshold is the best option with accuracy and precision above 94%. The development of these models shows that it is possible to develop an alternative solution to reduce operating costs compared to expensive commercial systems and to improve the usability compared to conventional methods of classifying road surfaces.