Deep Feature Retrieval With Human-in-the-Loop for Underwater Hull Fouling Segmentation

The inherent blurriness of underwater images, diversity of fouling patterns, and indistinct boundaries present significant challenges for underwater hull fouling segmentation tasks. To address these challenges, we propose a human–machine collaborative approach for underwater hull fouling segmentation, leveraging deep feature retrieval and local optimisation. Specifically, we first employ an image enhancement model as a preprocessing step to enhance underwater image clarity. Subsequently, a fine-grained segmentation model is utilised to generate initial segmentation results, which are then combined with a prior pixel label retrieval and propagation mechanism to identify locally optimised regions requiring refinement. Finally, manual correction of these localised regions is integrated with the segmentation model's predictions to achieve optimal segmentation performance. Experimental results on our self-constructed underwater hull fouling images dataset demonstrate the effectiveness of the proposed approach.