Identificação de defeitos do tipo “panela” em pavimento asfáltico por meio de redes neurais convolucionais.

Abstract

Road pavements contain surface defects that affect user’s comfort, safety and aspects such as cost and travel time. An alternative for improving highway management processes is the use of automated systems for obtaining and processing data using Artificial Intelligence, such as Convolutional Neural Networks (CNNs). The objective of this paper is to train CNNs to identify pothole-type defects, on roadways in Ceará. VGG16 and VGG19 networks were trained with high definition images and reached accuracy greater than 80%. The same networks, trained in high quality photos, were tested on images collected with smartphones, obtaining precision of 80% and accuracy of 40 and 60%, respectively. The results were encouraging and subsequent training with a greater number of images will allow the adjustment of a low-cost tool for automatic identification of surface pavement defects. Keywords: Neural Networks; Detection of Defects; Pavement Management; Automated System