• • Achieved >98% recall across 16 defect categories and a miss rate of only 2% for grade 4 (severe) defects, directly addressing the high missed-detection risk in manual CCTV inspection.
• • Improved per-segment detection efficiency by 425% over pure manual screening, significantly reducing labor costs and operational time for large-scale pipeline surveys.
• • The cross-frame annotation strategy (manual first/last frame labeling + interpolation + tracking) cuts annotation effort while maintaining consistency, enabling scalable dataset creation for deep learning models.
• • Integration of Focal Loss, defect classification/grading, and dynamic thresholding in a Faster R-CNN framework provides a robust mechanism to prioritize severe defects, enhancing practical utility for infrastructure maintenance.