Methodology for Training a Neural Network for Recognizing Road Pavement Defects

The paper describes a technology for analyzing the data of monitoring objects of the city's street and road network based on automatic detection and classification of the main types of road pavement damage (flatness, cracks, crack grids, potholes, repair maps, markings, and others) using convolutional neural networks. The results of a comparison of software tools used to create neural networks, as well as criteria for selecting libraries necessary for neural network training, are presented. The choice of the Google Colab cloud platform for neural network training is justified and a comparative analysis of its various versions is performed. The parameters of neural network training, statistics of image classification into three groups are given, and an algorithm for preparing input data is proposed. The values of IoU and F1 accuracy metrics by categories (classes) of objects are given. It is concluded that the amount of trained data, expressed as the number of pixels of regions of different classes rather than the number of those regions, is important for the quality of pixel classification.