Corrosion on ship hull surfaces poses a critical challenge in the maritime industry, directly affecting structural safety, fuel efficiency, and maintenance costs. Conventional manual visual inspection at drydocks is time-consuming, expensive, and poses potential safety hazards to inspectors. This study aims to develop an image-based ship hull corrosion detection system and objectively compare the performance of three deep learning architectures—YOLOv8, Faster R-CNN, and SSD300—under identical experimental conditions. The dataset combines the Marine Vessel Corrosion Dry-Dock Dataset (MaVeCoDD) with self-collected hull images, yielding a total of 76 images. Data were split consistently into training, validation, and testing subsets. Each model was trained via transfer learning from COCO pretrained weights and evaluated using mAP50, Precision, Recall, and F1-Score on the identical test set. Experimental results demonstrate that YOLOv8n outperforms the other architectures, achieving an mAP50 of 0.131, Precision of 0.313, Recall of 0.226, and F1-Score of 0.263. Faster R-CNN attained the highest Precision (0.651) but suffered from a low Recall (0.088), whereas SSD300 yielded the lowest performance with an mAP50 of 0.003. In conclusion, YOLOv8 proves to be the most suitable architecture for ship hull corrosion detection due to its superior balance between Precision and Recall compared to Faster R-CNN and SSD300.
Copyrights © 2026