earticle

영어

This study presents a U-Net-based deep learning segmentation model for automatic crack detection on concrete surfaces. Designed with a symmetric encoder-decoder architecture and skip connections, the model preserves spatial details and enables precise segmentation of irregular crack boundaries. Trained on a dataset of 40,000 images (20,000 positive, 20,000 negative), the model achieved an IoU of 0.484 and a Dice score of 0.513. It demonstrated strong real-time capability with 1,978.91 FPS, 7.18 GFLOPs, and only 1.87 million parameters. Visual results confirmed consistent segmentation across various crack types. The model's lightweight design and high efficiency support practical deployment in real-world applications such as drone inspections, CCTV systems, and edge computing. This approach highlights the potential of U-Net for scalable, real-time infrastructure monitoring.