Domain adapted probabilistic inspection using deep probabilistic segmentation

Abstract

This paper introduces the concept of domain-adapted probabilistic segmentation for marine vessel classification. The evolution of corrosion is continuous and it is, therefore, impossible to acquire marine vessel inspection datasets representative of the entire active fleet. Additionally, human surveyors introduce high levels of subjectiveness in the classification process, resulting in potentially multiple equally valid but ambiguous classification results. Consequently, deterministic inspection is flawed. The goal of this paper is to address these challenges by using a probabilistic approach to segmentation while performing domain adaptation to align the feature space across the different stages of age degradation. We test a Probabilistic U-Net on both simulated images and images from real vessels and compare it against two novel probabilistic models. We have evaluated the models using both quantitative — energy distance as distribution similarity — and qualitative — feature reduction visualization — approaches. Our results indicate that the combination of probabilistic segmentation and domain adaption could potentially have a high impact on marine vessel surveys in the future.