Framework for Crack Detection using Deep Learning

This study presents an innovative hierarchical classification approach aimed at detecting and classifying various types of road cracks. The methodology consists of three main stages: image preprocessing, binary classification, and hierarchical classification. The first experiment incorporates a binary classification model based on the OLenet architecture, where after preprocessing, the model is trained and tested on the Metu and ELpv datasets, achieving impressive accuracies of 97% and 92%, respectively. In the second experiment, the method amalgamates the Road Crack and Solar Cells datasets into a diverse dataset of 4,000 images. A Convolutional Neural Network (CNN) model is trained on this combined dataset using transfer learning techniques and architectures like VGG16, VGG19, and MobileNet, where the VGG19 model demonstrates superior performance with a classification accuracy of 98.69%. The final experiment involves a Mask RCNN model trained on the Road Crack dataset with 7,048 images of 13 distinct types of road cracks, offering the unique capability to classify multiple types of cracks within a single image and outperforming traditional CNN models with an accuracy of 90.42%. This novel approach holds significant potential for effective road crack detection and classification, contributing to enhanced road maintenance and safety. Future work will expand the hierarchical classifier with more crack types and further optimize the accuracy of all models. The main contributions of this study lie in its unique methodological approach, the effective use of diverse datasets, and the successful application of transfer learning and the Mask RCNN model.