Decoupled Contrastive Learning Constrained by Physical Feature for SAR Target Recognition

In the field of remote sensing target recognition, the fusion of synthetic aperture radar (SAR) and optical targets faces significant challenges due to the huge differences in feature representation. Current fusion recognition methods primarily focus on the feature alignment, overlooking the effective utilization of the distinct features inherent to each modality. A decoupled contrastive learning framework constrained by incoherent entropy (DCL-IE) is proposed to fuse the differential features of both SAR and optical modalities. DCL-IE can effectively enhance the model’s ability to distinguish between interclass differences between SAR and optical targets, thereby improving the accuracy of SAR target recognition. Specifically, decoupled contrastive learning (DCL) is designed to efficiently concentrate on different class features when oriented to cross-modal differential representations. The proposed relaxed label assignment algorithm can effectively distinguish between one specific class and the other classes, promoting the extension of DCL into the unsupervised learning domain. Furthermore, the physical incoherent entropy (IE) features are utilized to guide the learning direction of interclass representations, which enhances the extraction of intraclass features by leveraging frequency robustness. Extensive experiments with various target recognition methods on SAR and optical datasets, including FUSAR-Ship, FGSC-23, and FGSCR-42, demonstrate the effectiveness of the proposed framework.