Bayesian feedback-enhanced method in open set classification of ship-radiated sounds

Open set classification of ship-radiated sounds faces great challenges in detecting unknown ship types and maintaining recognition accuracy of known classes. Existing methods primarily focus on unknown ship-type detection but neglect semantic association between the novel ship-type and known ship-types. This paper proposed a Bayesian feedback-enhanced method that established mutual information exchange between classification and reconstruction pipelines. It introduced iterative probabilistic fusion where classification posteriors conditioned reconstruction error likelihoods through kernel density estimation, while reconstruction-based uncertainties refined classification decisions via feature fusion. Experiments on ShipsEar dataset demonstrated significant improvements, achieving 14% higher closed-set accuracy (84.1% vs 70.5%) and AUROC gain of 0.19 in unknown detection (0.924 vs 0.739) compared to the baseline. Notably, the method successfully associated 35.3% of unknown fish boats with semantically similar mussel boats, and 64.1% of ro-ro vessels with ocean liners. This advancement provides practical value to real-world maritime surveillance systems that typically work in an open world.