ABSTRACT
This research proposes a deep learning-based semantic segmentation method using U-Net to de-tect internal cracks in concrete cores extracted from in-service structures, employing X-ray Computed Tomog-raphy (CT). A major challenge in deep learning-based crack detection is the extreme class imbalance in the dataset. In the dataset used in this paper, the crack class accounts for only 0.5% and the void class for 0.7%, while coarse aggregates, mortar, and background dominate the composition. To address this imbalance, multi-class classification approach was adopted, and a novel Class-frequency-aware Focal Loss (CFL) was proposed. CFL applies nonlinear weighting according to the occurrence frequency of each class, thereby enhancing learn-ing for minority classes. In the results, multiclass classification outperformed binary classification, and the proposed CFL achieved notable improvements in F1-score. However, the high detection sensitivity of CFL was found to be accompanied by an increase in false positives in regions such as mortar areas.
