Training-Label Efficiency in Vision-Based Inspection of LNG Tank Cables: A Duplicate-Aware and Calibration-Controlled Benchmark of Supervised and Normal-Only Anomaly Detection

Vision-based sensing is increasingly important for liquefied natural gas (LNG) tank manufacturing, where missed defects in cable-related components can affect monitoring, control, and safety functions. However, industrial image datasets often contain duplicate or near-duplicate imagery, operating thresholds can be optimistically tuned on held-out test data, and abnormal labels are costly to obtain. This study evaluates LNG tank cable defect detection under a duplicate-aware and calibration-controlled protocol that separates model training, threshold calibration, and held-out testing. Using the AI-Hub LNG tank quality inspection image dataset, we compare ResNet-18, EfficientNet-B0, and ConvNeXt-Tiny supervised classifiers; PatchCore, PaDiM, and STFPM normal-only training anomaly detectors; and ResNet-18 training-label-budget experiments that retain all normal training images while varying the number of labeled abnormal training images used for model fitting. Across 36 de-duplicated final runs, EfficientNet-B0 achieved the strongest fully supervised performance (AUROC 0.9999±0.0001, F1 0.9956±0.0029). PatchCore was the best normal-only training anomaly detector (AUROC 0.8909±0.0109, F1 0.8666±0.0228). In the percentage-budget experiment, 1% labeled abnormal training samples already exceeded PatchCore in AUROC, F1, and false-negative rate under the fixed labeled-calibration protocol; in a finer 4–64 abnormal-training-label sweep with 90 runs, the first training budget exceeding PatchCore across AUROC, F1, and false-negative rate with 16 labeled abnormal training images. These results quantify abnormal training-label efficiency under a fixed labeled-calibration protocol, rather than the total abnormal annotation burden required to deploy the complete operating classifier.