Tri-State Prototype Self-Distillation for SAR Ocean Imagery Panoptic Segmentation

Synthetic aperture radar (SAR) ocean imagery panoptic segmentation provides a more all-round scene understanding than object detection and instance segmentation. However, it has not been extensively studied yet. We find that the stochastic queries of Transformer-based models from computer vision community lead to poor learning efficiency. Thus, we propose a novel Tri-State prototype self-distillation (TSP-SD) framework that replaces purely random query representations with Tri-State statistical Gaussian prototypes. Specifically, we decouple features into boundary, structure, and context via a Tri-State Gaussian prototype extractor (TGPE) to obtain hier-archical statistical prototypes for progressive evolution. Then, a cascaded statistical query injection (CSQI) decoder utilizes these proto-types to progressively guide the decoding process. Moreover, to reduce training uncertainty, a relative entropy alignment loss (REAL) is proposed to minimize the difference between the teacher and student. Experiments on SPSD and HRSID-PAN show that TSP-SD achieves 86.71% and 83.12% PQ, outperforming Mask2Former by 4.88% and 2.01%, respectively. Incorporating sea-land context further improves ship discrimination in complex coastal scenes.