Effect of Data Augmentation on Spalling Condition Classification using Deep Transfer Learning

Image processing is gaining attention in many structural research fields including condition assessment of concrete structural members using photos. on the other hand, the image processing field is completely revolutionized thanks to deep learning (DL). This research applies data augmentation besides deep learning technology to a civil engineering application, namely detecting the spalling condition of structural components from images. For this purpose, a dataset is used that was published by the structural image net project and only has a limited amount of images. Transfer Learning (TL) based on VGGNet (Visual Geometry Group) is presented and used on the mentioned dataset to minimize overfitting. Furthermore, a comprehensive data augmentation is also carried out. The algorithm boosted with data augmentation provides good recognition performance in spalling condition detection compared to training without data augmentation. These findings also point out the possible use of deep Learning in structural recognition tasks with limited data.

Keywords:

Deep Learning , Transfer Learning , Computer vision , Spalling condition , Data augmentation.