Binary classification of marine corrosion using deep learning: A comparative benchmark analysis across multi-source datasets

Marine corrosion presents a significant challenge to the durability and safety of maritime infrastructure, yet conventional inspection methods remain subjective, time-consuming, and limited in accuracy. This study benchmarks the binary classification performance of four state-of-the-art convolutional neural networks—VGG16, VGG19, InceptionV3, and ResNet50—across two independent experiments using distinct marine corrosion datasets. Experiment 1 employed 5105 images curated from three publicly available sources, while Experiment 2 utilized the complete 9,000-image Marine Corrosion Dataset. In both cases, images were labeled as corrosion_structures and healthy_structures, with no resizing or additional preprocessing applied, and training was performed in the KAI software framework using standardized augmentation, transfer learning from ImageNet weights, and consistent hyperparameter settings. Results demonstrated that ResNet50 achieved perfect accuracy (100 %) in Experiment 2 and exhibited strong generalization across datasets, while InceptionV3 and the VGG variants also delivered high precision, recall, and F1-scores above 96 %. The novelty of this study lies in combining two independent marine corrosion datasets for reproducible benchmarking, applying standardized CNN-based experiments with transfer learning, and establishing robust performance baselines for corrosion detection in maritime engineering. These findings highlight the feasibility of integrating deep learning into automated inspection workflows, enabling scalable, reliable, and efficient monitoring strategies that support predictive maintenance and extend the service life of marine assets.