A Hybrid Artificial Intelligence Framework for Risk-Oriented Port State Control Pre-Screening of Visual Ship Deficiencies

Port State Control (PSC) inspections are essential for maritime safety, but limited inspection resources make efficient vessel pre-screening increasingly important. This study investigates whether external vessel images can provide complementary information on visible ship deficiencies for PSC-oriented decision support. Five convolutional neural networks (ResNet18, GoogLeNet, MobileNetV2, DenseNet201 and SqueezeNet) were evaluated, followed by stacking and a hybrid framework combining ResNet18 deep features with handcrafted descriptors, mRMR feature selection, Principal Component Analysis, boosting classifiers and adaptive threshold optimisation. To make the comparison directly reproducible, duplicate image routes were removed, and all model families were evaluated on fixed target-specific 65/20/15 partitions. On the common independent tests, DenseNet201 provided the strongest balanced result for oxidation (balanced accuracy, 0.672; AUC, 0.742), whereas MobileNetV2 produced the largest point estimates for paint deterioration (balanced accuracy, 0.561; AUC, 0.659) and structural corrosion (balanced accuracy, 0.777; AUC, 0.885). Structural-corrosion recall was 3/4 = 0.750 (95% exact CI, 0.194–0.994) for MobileNetV2 and 2/4 = 0.500 (0.068–0.932) for the strict Hybrid PSC model, illustrating the uncertainty associated with rare positive cases. A controlled ablation further showed that changing only the operating threshold increased recall from 0.10 to 0.85 for oxidation and from 0 to 0.70 for paint deterioration, while increasing the alert burden. The results do not establish universal superiority of the hybrid representation; they show that representation and operating point should be interpreted jointly when visual AI is used as complementary PSC pre-screening support.