KFIA-Net: a knowledge fusion and imbalance-aware network for multi-category SAR ship detection

Multi-category synthetic aperture radar (SAR) ship detection is limited by heterogeneity in imaging mechanisms and severe class imbalance, yielding accurate localization but frequent misclassification. To address this issue, this paper proposes a Knowledge Fusion and Imbalance-Aware Network (KFIA-Net). Specifically, we first propose a Domain Knowledge Feature Extraction (DKFE) to extract and encode knowledge tokens from four priors. Second, a Knowledge Cross-Attention Fusion (KCAF) module is designed to perform interpretable and sparsely selectable channel modulation using cross-attention and FiLM decoding. Thirdly, we further design an Imbalance-Aware Loss Function (IALF) that combines prior calibration, minority category margin expansion, and knowledge-consistency weighting to reduce loss bias. Finally, systematic experiments and comparisons are conducted on three datasets: SRSDD-v1.0, FAIR-CSAR-v1.0, and NUDT-SARship-v1.0. Our KFIA-Net achieves mAP 50 scores of 64.29%, 37.99%, and 78.26%, and mAP 75 scores of 34.96%, 19.70%, and 66.36%, respectively. These results demonstrate knowledge injection simultaneously improves class accuracy and sustains robust localization. Furthermore, KFIA-Net requires only 11.47 M parameters and 66.79G FLOPs, achieving an inference speed of 47.21 FPS on a 1024 × 1024 input, achieving a good trade-off between accuracy and efficiency.