IDENTIFICATION OF ASPHALT PAVEMENTS DISTRESSES BY DIFFERENT DEEP-LEARNING ALGORITHMS

Abstract

The pavement condition’s assessment occurs mainly through laborious and time-consuming methods. Automating evaluations through computer vision can ensure greater safety for evaluators, efficiency, and productivity. The objective of this research is to compare different object detection architectures in the evaluation of distresses in urban flexible pavements. For this, the architectures You Only Look Once, Single Shot Detection, Faster Region-based Fully Convolutional Neural Networks and Faster Region-based Fully Convolutional Neural Networks were used to detect Pothole, longitudinal and transversal crack, alligator crack and patches. The analysis was performed by comparing the number of distresses detected correctly and incorrectly. The Single Shot Detection architecture was the most accurate in identifying the deterioration analyzed, with 286 distresses correctly identified. The You Only Look Once model was the second most efficient, with 77 distresses identified correctly. The Region Proposal Network algorithm, present in the Faster Region-based Fully Convolutional Neural Networks and Faster Region-based Fully Convolutional Neural Networks models, was unable to learn to generalize regions of interest from the images, which hindered the identification of deteriorations. It was concluded that the use of deep learning algorithms to identify distresses in flexible pavements is feasible, with efficient architectural models, capable of identifying different deterioration types.