Dual-Modal Representation Interactive Offset Learning Method for Maritime Radar Target Detection

Abstract

It is one of the important means to enhance maritime target detection in sea clutter by characterizing the differences between targets and clutter using a multimodal manner. Therefore, this letter proposes a dual-modal learning-based marine target detection framework called dual-modal representation interactive offset learning (DMRIOL) method. To be specific, we first construct two modalities of radar echoes: smoothed pseudo Wigner–Ville distribution-based time–frequency (SPWVDTF) representation and symmetric dot pattern-based phase information (SDPPI) representation, to characterize the echo signals from different domains. Then, to fully learn the complementary characteristics between the two modalities and eliminate redundancy, we develop an information communication module (ICM), which is embedded in the dual-modal feature extraction layers. Subsequently, we propose a multihead attention-based cross-modal fusion strategy (CMFS) to achieve feature alignment and fusion. Experiments conducted on the public IPIX1993 and SDRDSP datasets demonstrate that the proposed method achieves higher target detection probability compared with some state-of-the-art detectors.