Physics-based Generative Augmentation Pipeline for Fine-Grained Ship Classification
作者
Bo Liu,Chenhui Zhu,Yinan Xing,Yunfeng Zhang
出处
期刊: [Institute of Electrical and Electronics Engineers] 日期:2026-08-01卷期号:: 1514-1518
标识
DOI:10.1109/igarss60577.2026.11730551
摘要
Data scarcity and high inter-class similarity severely hinder fine-grained ship classification in remote sensing. While physics-based simulation provides a scalable data source, it suffers from a significant discrepancy between synthetic and real domains. Conventional global style transfer methods fail to bridge this gap, often inducing structural distortions and texture artifacts by treating the rigid ship foreground and dynamic ocean background uniformly. To address this, we propose a dual-stage pipeline comprising a Physics-based Visible Light Image Simulation System (PVLISS) and an Instance-Aware Style Transfer Network (IAST-Net). Specifically, IAST-Net exploits the privileged pixel-wise masks inherent in the simulation process to explicitly disentangle the ship from the background. This ensures realistic sea surface rendering while preserving the geometric integrity of the vessels. Experiments demonstrate that our pipeline generates high-fidelity images and significantly improves model generalization, particularly for rare ship classes, outperforming state-of-the-art augmentation methods.