Detecção de anomalias em pavimentos asfálticos usando redes neurais para segmentação de imagens

Abstract

Pothole identification is a crucial step in defining road maintenance and intervention strategies. By detecting road defects, transportation authorities can prioritize repairs and allocate resources more effectively, resulting in reduced operating costs and improved road safety. However, conventional manual detection methods are time-consuming and costly due to the high consumption of human and material resources. Given this scenario, the use of digital image processing and deep learning concepts to identify potholes has proven to be a promising approach to improve maintenance and intervention strategies on highways. In this study, an automatic method was developed based on advanced image segmentation techniques, based on specialized literature. The method uses the U-NET network as a base, replacing the learning step with an EfficientNetV2 encoder, which allows a more efficient extraction and reconstruction of the images. The speed and efficiency of this method in detecting anomalies on pavement surfaces stands out. In addition, the method was compared with other existing models, showing promising results during the tests performed. A highlight is the low number of parameters used by the network, totaling only 5.4 million, compared to other models that ranged between 14 and 45 million parameters. This demonstrates the efficiency and ability to obtain satisfactory results with a reduced number of parameters.