Evaluating Generative Image Expansion for Long-Range Maritime Vision Tasks

Abstract

Synthetic image generation is increasingly used to augment visual datasets when real-world data is limited or difficult to capture. However, generative techniques do not simply extend existing images; they actively construct contextual assumptions about background continuity, spatial relationships, and scene structure. In decision-relevant settings, these assumptions can obscure uncertainty and introduce ambiguity that affects both model behavior and human interpretation. This paper examines the use of generative image expansion to induce distance-related perceptual stress in naval vessel imagery, motivated by the needs of maritime decision support under conditions aligned with Tactical Decision-Making Under Stress (TADMUS). Rather than evaluating downstream model performance, we focus on the interpretability and contextual integrity of augmented images as perceived by human annotators. We construct a dataset of perceptually degraded image variants using multiple generative platforms and assess output quality using a structured, context-focused annotation protocol.