The Effectiveness of Data Augmentation Methods in Multi-Defect Classifications from Façade Images Using Transfer Learning

Falls from façades due to defects pose safety threats to public and require regular inspections. Conventional inspection methods are manual and based on the expertise of inspectors resulting in undetected defects and subsequent incidents and accidents. Opportunities that enable vision-based identification of defects, such as deep learning (DL), are available but require abundant labelled images for robust solutions. Yet, data collection and labelling for domain-specific tasks are expensive and time-consuming, resulting in limited training data and/or imbalanced datasets. Previous studies have successfully employed various DL architectures to increase model accuracies for detecting façade defects but were observed to be limited due to imbalanced and/or small datasets. The aim of this study is to mitigate the problem of data scarcity by deploying various combinations of data augmentation techniques and evaluating the accuracy of models developed using the augmented data produced by these techniques. We applied transfer learning using Mask R-CNN and incorporated two novel data augmentation approaches (CutMix and MixUp) along with traditional techniques such as geometric transformations. The accuracies of models in multi-defect detection are evaluated.