The Near-Field SAR Single-Image Generation Method Based on GCS-ConSinGAN

To address the challenges of limited sample availability and insufficient structural fidelity of existing single-image generation models for near-field synthetic aperture radar (SAR) imagery in active millimeter-wave security screening scenarios, a ConSinGAN model integrating a Grouped Attention Gate (GAG) and a Channel-Spatial Attention Module (CSAM), termed GCS-ConSinGAN, is proposed. Built upon the parallel multi-scale architecture of ConSinGAN, the proposed model introduces GAG into the cross-scale connections of the generator to selectively propagate features and suppress the accumulation of pseudo-textures. In addition, CSAM is embedded at the front end of the discriminator to jointly model channel dependencies and spatial correlations, thereby enhancing the discrimination capability for critical scattering structures. Experimental results on a self-constructed near-field SAR dataset demonstrate that the proposed GCS-ConSinGAN reduces the Single Image Fréchet Inception Distance (SIFID) from 0.46 to 0.28, achieving a reduction of 39.13%, while requiring only 21 min of training time. Furthermore, when the generated samples are used for training set augmentation, the classification accuracies of InceptionV3, ResNet101, and Vision Transformer (ViT) in automatic target recognition (ATR) tasks are improved by 0.32%, 0.42%, and 0.77%, respectively. Ablation studies further reveal that GAG and CSAM contribute 23.9% and 15.22% of the SIFID reduction, respectively, demonstrating a synergistic effect between the two modules.