Zero-Shot Sonar Image Quality Assessment via Contour Semantic Shift

Underwater sonar imaging is a key technology in modern ocean exploration and maritime defense. However, distortions often occur during transmission or compression due to the harsh underwater communication environment. These distortions can severely degrade image quality. Therefore, implementing Sonar Image Quality Assessment (SIQA) of transmitted images is crucial to ensure their reliable performance in downstream tasks. In practice, collecting large-scale annotated sonar datasets is both costly and challenging, and high-quality reference images are often unavailable. Therefore, SIQA methods must be reference-free and independent of training data. To address this challenge, we introduce a novel zero-shot SIQA paradigm, where task-oriented image quality is defined by the magnitude of semantic shift under contour degradation. High-quality images, with clearer contours, exhibit larger semantic shifts when perturbed; in contrast, low-quality images with blurred structures show smaller shifts. We simulate contour degradation by filtering the mid-frequency components and quantify the resulting semantic shift using cosine similarity. This approach bypasses conventional feature engineering and regression-based modeling, directly using the measured semantic shift as the quality score, thereby enabling zero-shot evaluation. Experiments on sonar datasets show that our method outperforms existing zero-shot IQA baselines and demonstrates strong potential for practical application.