A novel application of attention U-Net for marine biofouling segmentation

Abstract Marine biofouling presents significant challenges in the maritime industry, including increased drag, fuel consumption, and maintenance costs. Traditional inspection and mitigation methods are labour-intensive and time-consuming, highlighting the need for automated approaches of biofouling detection and analysis. This study aims to bridge the existing literature gap by introducing an enhanced Attention U-Net architecture specifically optimised for the semantic segmentation of marine biofouling in real-world conditions. Our model incorporates spatial attention gates within the skip connections and squeeze-and-excitation modules within each convolution block of a traditional U-Net framework. It was trained and tested on an annotated dataset of 504 in-water biofouling imagery collected from multiple ship hull surveys, provided by diving companies, classification societies, and NTUA's archive. It contains images captured under various environmental conditions, which enables better model generalisation. The aforementioned pipeline resulted in a validation Dice coefficient of 0.814 and a macro-accuracy of 0.689, suggesting advanced segmentation capabilities. Promising implications arise for deployment in automated inspection systems, potentially enhancing the efficiency of hull and offshore structure inspections by reducing manual effort and improving detection accuracy.