Machine learning techniques provide great opportunities to conduct research on a wide range of topics, mainly to solve complex problems. This project’s primary motivation is to detect concrete pavement structure damage automatically. For this purpose, extensive research on neural network topics is carried out. Then, computational image-processing techniques are implemented.
A data set of images from national roads in Costa Rica provided by the National Laboratory of Structural Materials and Models of the University of Costa Rica (LanammeUCR) is used. The images were divided into two groups, one for training and another one for testing. Two types of damage, cracks, and patches, were initially analyzed, and then the study was expanded to three damages. Masks with damage labels were created for the total number of images to use in the training and testing of the network. The masks were filtered to ensure that they only contained the colors associated with the labels.
The associated hyperparameters were investigated and defined based on studies or by performing iterative processes. Different Neural network (NN ) models were trained and tested, those were compared. Then, the dataset was increased with online and offline augmentation and the results of the implementation with the ExtremeC3Net, UNetResNet, and UPerNet models were compared. Promising results of Intersection over Union (IoU ) values major to 0.5 were obtained from the UPerNet and UNetResNet with a 34 backbone.
Moreover, the same models were implemented with a dataset in which three damages were labeled. Lower mean IoU values were obtained; however, the results were considered satisfactory for the number of images and masks used in training. Ideas for the further development of this project are presented at the end of this document.