Occlusion SAR Target Recognition Based on Learnable Fractional Gabor Transform and Local Scattering Extraction Network

Aiming at the problem of significant performance degradation of synthetic aperture radar (SAR) image recognition network under target occlusion conditions, this letter proposes a learnable fractional Gabor transform and local scattering extraction network (FGT-LSENet) for occluded SAR target recognition. In the proposed method, we first design the learnable Gabor transform module to guide the network to perform global information extraction of features with multilevel fusion. Subsequently, the local scattering module is utilized to extract the strong scattering features that occlude the target key structure. Finally, the convolutional, global, and local scattering information are fused and the network is optimized using triplet loss function and central loss function. Experiments with different degrees of occlusion on the moving and stationary target acquisition recognition (MSTAR) dataset show that the proposed method achieves higher recognition rates with different degrees of occlusion compared to existing network models.