SAR target recognition based on hierarchical azimuth aware feature enhancement network

Synthetic aperture radar (SAR) target recognition faces critical challenges due to the azimuth sensitivity of the imaging mechanism, which induces significant intra-class variance. This issue is further exacerbated under extended operating conditions (EOCs), such as noise, occlusion and depression variations etc. severely degrade recognition performance. To address these limitations, this paper proposes a hierarchical azimuth aware feature enhancement network (HAAFENet). Distinct from conventional methods, HAAFENet explicitly embeds azimuth priors into the deep learning framework to extract robust, orientation-invariant representations. Specifically, a novel azimuth aware feature enhancement module (AAFEM) is designed to dynamically generate parameters for adaptive channel and spatial modulation based on the target’s orientation. To preserve feature representativeness, a hierarchical feature fusion module (HFFM) is incorporated to integrate deep semantic features with shallow geometric details. Furthermore, we introduce two synergistic loss functions to optimize the feature space: (1) an azimuth difference feature consistency loss, which utilizes contrastive learning to enforce intra-class feature similarity across large angular variations; (2) a dynamic angular margin loss, which adjusts classification boundaries based on the equivalent number of looks (ENL) to rigorously constrain images corrupted by speckle noise. Extensive experiments on the MSTAR and SARSIM datasets demonstrate that HAAFENet achieves excellent performance, exhibiting superior robustness and discriminability under both standard and diverse extended operating condition.