FPTF-YOLO: Fully Polarimetric Time–Frequency YOLO for Maritime Radar Weak Target Detection and Localization

Traditional maritime radar target detection methods primarily rely on handcrafted features and single-polarization data, which limits their ability to suppress sea clutter false alarms and enhance weak target detection. To address these limitations, this paper investigates maritime weak target detection and localization using fully polarimetric radar echo data. First, a Fully-Polarimetric Time-Frequency YOLO (FPTF-YOLO) network is developed based on the YOLOv11 framework to detect and locate maritime weak targets. The proposed FPTF-YOLO employs full-polarization time-frequency inputs through four parallel backbone branches corresponding to HH, HV, VH, and VV channels, and designs a polarization feature fusion (PFF) stage that integrates a simple channel attention (SCA) module to enhance multi-scale feature integration. In addition, squeeze-and-excitation (SE) and coordinate attention (CoordAtt) modules are incorporated into the Neck stage to further refine discriminative feature representations. Then, a fully polarimetric normalized time-frequency (TF) dataset is constructed using measured IPIX radar data and simulated data, comprising training, testing, and an independent unseen IPIX test dataset that is excluded from training procedure to assess generalization performance. Experimental results on the constructed datasets demonstrate that the proposed FPTF-YOLO outperforms its competitors, and achieves average detection probabilities of 0.923 and 0.932 for observation times of 512 ms and 1024 ms, respectively. Moreover, the FPTF-YOLO exhibits stable detection performance on independent unseen IPIX test dataset without retraining, demonstrating reliable generalization capability.