ResNet neural network hyperparameter tuning for Rigid Pavement Failure Assessment

Abstract

Rigid pavement roads do not have adequate maintenance, since the inspection stage is carried out “manually”, which is not reliable or efficient, as well as requiring a greater amount of labor, time, and high cost. To solve the problem, it is proposed to evaluate the rigid pavement condition using ResNet neural networks with images obtained through a conventional 2D camera. The objective of the work was to recognize three types of failures in the rigid pavement: joint peeling, corner peeling, and corner crack. For the preprocessing phase I use image normalization and resizing, the number of images was increased through geometric transformations by 12.21%. A convolutional neural network of ResNet-18 type architecture was used. As a learning transfer technique, model tuning was used, since we not only changed the output network, but also the hyperparameters of the convolutional layers. The contribution of the present work was the refinement of the hyperparameters for the modification of the ResNet-18 neural network taking into account the iteration in the learning rate that goes from 1e-4 to 1e-2. The results were: accuracy 88.73%, sensitivity 81.63%, a specificity of 92.47%, the precision of 85.10%, and finally an F1 score of 83.33%. Three of the model’s evaluation indices have values higher than 0.71 while the fourth has a value of 0.55, which indicates that there will be a good performance with the proposed model. This work can be improved by increasing the number of images or by making a hybrid model.