A novel dual-polarization SAR ship classification method based on a polarimetric attention-guided and frequency domain feature fusion

This is the authors' abstract. We don't add key points for this paper.

Owing to the capabilities of all-weather and day-night, Synthetic Aperture Radar (SAR) ship classification serves as a critical component in maritime surveillance. Deep-learning methods offer state-of-the-art performance in the ship classification task, yet make full use of dual-polarization data. In this work, a Polarimetric Attention-Guided Network (PAG-Net) is proposed for dual-polarization SAR ship classification. To fully exploit dual-polarization information, a tri-branch feature extraction backbone is constructed by jointly executing Polarimetric Feature Guided Attention (PFGA) modules and Feature Interaction Extraction (FIE) modules. In each branch, the PFGA and FIE are collaboratively designed to concentrate on the key feature through cross-polarization attention mechanism. Furthermore, a frequency-context up-sampler is built based on the Frequency Domain Feature Fusion and Enhancement (FDFE) module, which can refine feature maps and reduce computational overhead. Experiments are conducted on a balanced OpenSARShip dataset for 3-class and 6-class classification tasks. The results show that the proposed PAG-Net method outperforms the known deep learning-based methods such as ResNet50, DenseNet, YoloV11, SwinTransformer V2 and ConvNeXt V2 in terms of accuracy, precision, recall and F1 score.