MaOutCNN: A physical mechanism coupled MambaOut-CNN detection network with maiden-released interference polarized ship detection dataset

Unlike traditional ship detection, polarimetric synthetic aperture radar (PolSAR) intelligent ship detection faces the following critical challenges: inaccurate revelation of polarimetric responses, insufficient physical interpretable network and lack of consideration of complex electromagnetic environment. To address these issues, this paper proposes a physical mechanism coupled MambaOut-CNN network for PolSAR ship detection under complex interference scenarios. First , by arduously collecting six types of spaceborne and airborne real PolSAR data covering three mainstream frequency bands, this paper releases an interference polarized ship detection (IPSD) dataset for the first time, wherein four types of typical interferences are involved and a total of 934 samples are presented. Second , a regularized systematic polarimetric decomposition (RSPD) framework for ship targets is proposed. On the basis of introducing rotated dihedral and planar resonator scattering models unique to ships, by defining the hierarchical absolute dominant scattering principle and designing a simultaneous model inversion strategy, this decomposition framework achieves fine characterization of ship target scattering in terms of emphasis of its exclusive scattering. Finally , with physical prior of the RSPD, a physical mechanism coupled MaOutCNN ship detection network by designing CNN and MambaOut dual-branch inputs is proposed, which integrates multi-scale parallel, multi-branch convolution and dynamic feature fusion modules, and introduces a collaborative perception loss function. This guarantees a generalized and reliable detection of dense targets in near-port areas and widespread targets in open seas. Experimental results show that the proposed RSPD accurately reveals and outlines prominent the global and local structural scattering of ship targets. Moreover, compared with 12 advanced ship detection networks, MaOutCNN achieves the optimal detection performance on both the IPSD and non-interference datasets, with average detection accuracies of 89.7% and 94.2% respectively, which are 3.6% and 3.1% higher than the baseline model. In addition, the results of ablation experiments and heatmap visualization are also discussed. The dataset and benchmark are publicly available at https://github.com/zzandcc/ipsd.git .