Toward Improved Surface Defect Segmentation in Concrete Structures: Enhancing U-Net for Multi-Class Segmentation

Detecting surface defects is essential for evaluating the condition and structural integrity of concrete infrastructure. While recent research has applied deep learning for automated defect detection, most efforts focus solely on cracks. A smaller subset addresses both cracks and spalling—two primary defect types—but multi-class detection remains challenging. Traditional object detection methods use bounding boxes for localization, whereas semantic segmentation offers pixel-level precision, enabling more accurate defect quantification. This study employs a modified U-Net, a semantic segmentation model, to classify concrete surfaces into three categories: cracks, spalling, and non-defective regions. The model’s performance is compared to the original U-Net in terms of accuracy and computational efficiency. Results show that the modified U-Net achieves a mean pixel accuracy of 75.2%, a mean intersection over union (IoU) of 66.5%, and a mean Dice coefficient of 78.6%, demonstrating its effectiveness for detailed multi-class defect segmentation.