Abstract
Offshore wind farm (OWF) inspections are subject to many challenges both physical and contextual, footage is often noisy and full of naturally salient objects with little to no relevance to the inspection. One of the key challenges of these inspections is effectively finding multiple/similar frames in the footage that show defects. This is because conventional embedding methods might focus too heavily on irrelevant but salient objects. In this research I conduct a pilot user study to capture the eye movements of experts and non experts (in the area of OWF inspections) while viewing underwater footage. This data was then used to train a model to generate salient masks mimicking expert and non expert gaze patterns. The model successfully replicated gaze locations and expert polarity based on the data from only two experts and five non experts. The model showed evidence of identifying relevant and irrelevant regions but more data will be needed to reliably judge the semantic relevancy of objects, hence this work serves as the foundation for a future study.