Target Recognition of Few-sample SAR Images Based on Two-step Domain Adaptation

Convolutional neural networks (CNNs) have achieved widespread applications in image recognition domain due to their superior performances. However, in the context of synthetic aperture radar (SAR) image-based ground vehicle target recognition, CNN models are prone to overfitting during the training phases due to the lacks of large-scale real-measured datasets. While SAR images based on target scattering characteristics, obtained through electromagnetic simulation, are relatively easier to acquire, there exists a significant domain discrepancies between simulated and real-measured images, therefore, directly using simulated images for training leads to domain shifts, which severely impacts recognition performances. To address this issue, an algorithm that utilizes electromagnetic simulation SAR images for auxiliary training is proposed in this paper. The approach employs style transfer to reduce the visual differences between simulated and real images and achieve feature alignment through adversarial learning. Domain adaptation is implemented progressively at both the image pixel and feature levels, enabling the feature extraction network to capture common target classification features from both domains. Experimental results on a typical SAR vehicle target dataset demonstrate that when the ratio of real-measured to simulated training data reaches 3 ∶ 10, the recognition accuracy of the proposed method achieves approximately 95%, outperforming classical domain adaptation algorithms and significantly enhancing the generalization capabilities of SAR target recognition models.