Abstract
Road surface integrity is critical for ensuring vehicular safety and reducing maintenance costs, yet traditional visual inspections are laborious and subject to human error. This study presents a tailored 2D U-Net architecture for automated pixel-wise segmentation of pavement damage specifically cracks and potholes from smartphone-captured RGB imagery. The proposed model employs an encoder–decoder design with symmetric skip connections to merge high-level contextual information and fine-grained spatial features. A weighted cross-entropy loss function is used to mitigate severe class imbalance between sparse damage pixels and dominant intact surface areas. Training on a curated dataset of 1,000 labeled road images demonstrates steady convergence: training loss decreases from 2.65 to 2.25, while validation accuracy rises from 65 % to 76 % over 36 epochs. Qualitative results reveal that coarse crack clusters are reliably detected by epoch 20, with finer details emerging by epoch 30. Mobile deployment tests on a mid-range smartphone yield an average inference time of 45 ms per frame, and TensorFlow Lite quantization reduces model size by 30 % with less than 1 % IoU degradation. These findings confirm that the 2D U-Net framework balances segmentation accuracy and on-device efficiency. Future work will explore hybrid loss formulations and expanded datasets including nighttime and adverse weather conditions to further enhance generalization and robustness.