Dual-View Screening of Ship-Coating Surface Defects Using Frozen Visual Representations and Local Evidence Aggregation

Shipyard coating images are affected by reflections, surface texture, structural edges and uneven defect-class distributions. This study evaluates a practical point-level screening framework through offline analysis of dual-view coating images from a single shipyard. We feed frozen DINOv2-S/14 features to a 385-parameter linear patch head with normalised log-sum-exp aggregation. Training weights balance inspection points, separate normal calibration points determine thresholds, and the higher score of two assessment views triggers an alert. The model was fitted on 39 points and frozen before acquisition of the 146 evaluation points. It detected 72 of 73 sagging or wrinkling points (98.6%), with nine false alarms among 65 normal points (13.8%). Compared with matched mean pooling, it detected four more known-defect points and produced three fewer normal false alarms. It also detected six of eight other-defect points. Adding the second view detected one more known-defect point and one other-defect point, while referring four more normal points for review. In the original evaluation, median processing time was 0.64 s per JPEG pair. An additional 25-point field study with two photographers and two blinded region-of-interest (ROI) operators retained all 18 defect detections in each first-round series, with 0–2 false alarms among seven normal points. Dual-view decisions remained correct at all eight reacquired points. The approach provides point-linked visual records and screening results to support expert review and reinspection.