2026 Volume 7 Issue 3 Pages 346-353
This study conducted a fundamental investigation of an image analysis method using deep learning to improve the efficiency of screening frost-damaged prestressed concrete sleepers and to contribute to quantitative soundness evaluation based on scaling area. A multi-class semantic segmentation model was developed by utilizing an existing U-Net model for crack detection, while maintaining its crack extraction capability and extending it to detect scaling. The validation results showed that crack extraction still had issues in extracting small regions and crack tips, and cracks were missed in some images. However, the model showed certain applicability to primary screening. For scaling, the proposed model showed good screening performance, and its region extraction performance indicated the potential for estimating scaling area.