Consistency-Regularized GAN for Few-Shot SAR Target Recognition

TL;DR

This work tackles the limited-label problem in SAR target recognition by introducing Cr-GAN, a consistency-regularized GAN with a dual-branch discriminator that decouples adversarial training from representation learning. It employs channel-wise feature interpolation and dual-domain cycle consistency to generate diverse, high-fidelity SAR samples while avoiding data augmentation leakage, enabling effective self-supervised pretraining (SimCLR) on synthetic data and robust fine-tuning on scarce real labels. Across MSTAR and SRSDD, Cr-GAN achieves state-of-the-art few-shot accuracies (e.g., $8$-shot: $71.21\%$ on MSTAR and $51.64\%$ on SRSDD) with an order of magnitude lower parameter and computational footprint than diffusion models. The results demonstrate improved data efficiency, stable training in extreme data regimes, and meaningful downstream feature representations, suggesting Cr-GAN as a practical regularization framework for SAR and potentially other domains with severe data scarcity.

Abstract

Few-shot recognition in synthetic aperture radar (SAR) imagery remains a critical bottleneck for real-world applications due to extreme data scarcity. A promising strategy involves synthesizing a large dataset with a generative adversarial network (GAN), pre-training a model via self-supervised learning (SSL), and then fine-tuning on the few labeled samples. However, this approach faces a fundamental paradox: conventional GANs themselves require abundant data for stable training, contradicting the premise of few-shot learning. To resolve this, we propose the consistency-regularized generative adversarial network (Cr-GAN), a novel framework designed to synthesize diverse, high-fidelity samples even when trained under these severe data limitations. Cr-GAN introduces a dual-branch discriminator that decouples adversarial training from representation learning. This architecture enables a channel-wise feature interpolation strategy to create novel latent features, complemented by a dual-domain cycle consistency mechanism that ensures semantic integrity. Our Cr-GAN framework is adaptable to various GAN architectures, and its synthesized data effectively boosts multiple SSL algorithms. Extensive experiments on the MSTAR and SRSDD datasets validate our approach, with Cr-GAN achieving a highly competitive accuracy of 71.21% and 51.64%, respectively, in the 8-shot setting, significantly outperforming leading baselines, while requiring only ~5 of the parameters of state-of-the-art diffusion models. Code is available at: https://github.com/yikuizhai/Cr-GAN.

Figures (16)

• Figure 1: (a) The overall architecture, illustrating the data flow for both image reconstruction and feature interpolation. A real image is encoded into a feature, which can be either reconstructed directly or interpolated with other features from a memory bank to synthesize novel images. (b) Conceptual illustration of the Image reconstruction, which enforces that a reconstructed image should be close to its real-world counterpart. (c) Conceptual illustration of the Feature reconstruction, which regularizes the feature space by encouraging meaningful representations when interpolating between features.
• Figure 2: The proposed pipeline, built upon Cr-GAN and self-supervised learning, comprises three key stages: (1) Cr-GAN Generation Method: We first train our Cr-GAN model on the limited real data. The objective is to learn the underlying data distribution and synthesize a large-scale, diverse dataset for the subsequent pre-training stage. (2) Self-Supervised Learning: We leverage the synthetic dataset to pre-train a backbone network. By employing a contrastive learning objective, the model learns robust and generalizable feature representations without requiring any labels. (3) Downstream Task: The pre-trained backbone is fine-tuned on the original, scarce set of labeled images, adapting the learned representations to the specific downstream classification task.
• Figure 3: The proposed Cr-GAN framework. The discriminator's feature branch encodes real images into latent features. Channel-wise interpolation then synthesizes mixed features. The generator uses both original and mixed codes to produce reconstructed and novel synthetic images, respectively. Two losses provide supervision: the image reconstruction ensures fidelity, and the feature reconstruction aligns the generated features with the mixed features to ensure consistency. The superscripts $r, m, g$ denote real, mixed, and generated data.
• Figure 4: Two-branch discriminator architecture. A shared convolutional body extracts features that are fed into two distinct heads: an adversarial head for real/fake classification and a feature head that outputs the mean and variance.
• Figure 5: Illustration of shortcut learning. (a) When synthesizing a new feature by interpolating two existing ones, a naive model might learn a trivial mapping instead of generating a semantically novel representation. (b) A single-pixel variation in the image space can correspond to a substantially amplified discrepancy in the feature space, indicating a non-robust representation.