Multi-granularity feature enhancement for robust remote sensing ship classification

Ship classification plays a critical role in civilian applications such as environmental monitoring and maritime navigation. While existing methods achieve high accuracy with high-resolution images captured under ideal conditions, their performance deteriorates significantly in uncontrolled environments with uncertain data quality. This degradation poses two key challenges: the loss of structural details in low-quality images leads to semantic information loss during feature extraction, and the reduced discriminative power of deep representations further limits classification robustness. To address these issues, we propose the Multi-Granularity Feature Enhancement Network (MGFE-Net) to enhance the robustness of ship classification. First, to mitigate the information loss during features extraction, we introduce a target awareness module that preserve semantic information in intermediate features layers while filtering out irrelevant noise. Second, to enhance features discriminability, we design an adaptive feature enhancement module that selectively amplifies the most discriminative features withinmulti-scale deep representations. Finally, to fully leverage the complementary strengths of semantic and discriminative features, we employ multi-granularity feature fusion strategy that integrates shallow and deep-layer features to improve ship classification performance. Experiments on ship datasets demonstrate that the proposed MGFE-Net significantly outperforms recent state-of-the-art methods.