CTRS : An enhanced YOLOv10 ‐based instance segmentation model for multi‐key block rough‐surface inspection in precast concrete components

Abstract Accurate inspection of multi‐key‐block rough surfaces is important for the quality control of precast concrete components, but manual assessment remains subjective and inefficient. This study constructed MBsurface1600, comprising 1600 images derived from 400 independent physical rough surfaces, and developed CTRS, an enhanced YOLOv10‐based instance segmentation model incorporating C2f‐Faster, C3FNet–AFR, and SegNextAttention. Across five repeated runs, CTRS achieved 86.43% ± 0.46% mAP 50 and 61.14% ± 0.47% mAP 50‐95 , exceeding the YOLOv10‐based baseline by 5.15 and 4.77 percentage points, respectively. CTRS also achieved 367.0 ± 2.9 FPS, exceeding the baseline by 13.2 FPS while reducing the parameter count and computational complexity by 8.9% and 10.9%, respectively. Evaluation on 40 untouched test images yielded 92.5% pass/fail decision accuracy and a mean absolute qualification‐rate error of 2.8 percentage points.