• Developed a ShuffleBlock-based Mask R-CNN for ship hull defect detection. • Introduced Texture-Aware Modules to capture corrosion-specific irregularities. • Integrated attention mechanisms for robust multi-scale feature refinement. • Achieved IoU of 81.31% (0.5) and 89.74% (0.75) on Roboflow Corrosion dataset. • Outperformed YOLOv8, U-Net, Mask R-CNN, and Faster R-CNN in accuracy and robustness. Ship hull defect detection is crucial for early identification of corrosion and damage, enhancing structural safety, preventing costly failures, and boosting maritime operational efficiency. Due to prolonged exposure to harsh marine environments, ship hulls age over time and are prone to corrosion, rust, and surface defects. This significantly weakens their structural integrity and reduces the ship's operational efficiency. To enable early stage defect detection, a deep learning-based computer vision architecture is proposed for automated ship hull inspection. The model accurately detects defects and performs instance segmentation of corrosion on hull surfaces. It extends the Mask R-CNN architecture with three key techniques: a lightweight ShuffleBlock-based backbone for efficient feature extraction, Texture-Aware Modules to capture corrosion-specific irregularities, and attention mechanisms for enhanced multi-scale feature refinement. Trained on the Roboflow Corrosion dataset, the proposed network demonstrates strong performance in detecting the hull defects, while also providing reliable confidence measures through qualitative evaluation. Quantitative results show that the model achieves average IoU scores of 81.31% at a 0.5 threshold and 89.74% at a 0.75 threshold, with corresponding true positive rates of 83.87% and 54.84%, respectively. These results highlight the effectiveness of the proposed approach as a tool for real-time maritime corrosion assessment and maintenance planning.
Pravin et al. (Wed,) studied this question.