SAR Target Recognition With Image Generation and Azimuth Angle Feature Constraints

Abstract

Under high-resolution synthetic aperture radar (SAR) observation, SAR targets exhibit significant azimuth sensitivity, with their scattering characteristics undergoing nonlinear changes with observation angle (manifesting as notable intra-class differences for the same type of target, while different classes present similar scattering features within the same azimuth angle range). This azimuth limitation of the data results in the recognition model struggling to effectively decouple azimuthsensitive features, leading to inter-class confusion at specific observation angles. To address this issue, this paper proposes a multi-azimuth SAR vehicle image generation method based on azimuth angle feature (AAF) constraints. The method generates SAR images at multiple azimuth angles, thereby enhancing the performance of SAR automatic target recognition (ATR). Considering the complexities of SAR imaging, our method adopts a coarse-to-fine generation strategy. In the coarse stage, AAFs are extracted from measured SAR images and intermediate AAFs are generated using a generative adversarial network (GAN) framework. In the fine stage, image-level generation is achieved via AAF constraints, producing high-quality target SAR images at intermediate azimuth angles. Unlike traditional methods based on one-dimensional angle labels, our approach utilizes twodimensional AAFs as input. It reduces the difficulty of learning azimuth angle features and enhances the physical consistency of the generated SAR images. Through the coarse-to-fine generation strategy, our method makes full use of AAFs. This approach reduces dependence on large datasets, enhancing its suitability for generating vehicle target images in data-scarce scenarios. Additionally, by integrating pixel-level loss, structural loss, and wasserstein loss, the quality and stability of the generated images are further optimized. Experimental results demonstrate that our proposed method significantly outperforms existing approaches. It achieves improvements for SAR image generation in terms of structural features, scattering characteristics, and deep features. In ATR tasks, the generated data was incorporated into both the Moving and Stationary Target Acquisition and Recognition (MSTAR) dataset and a self-collected dataset. The target recognition model was retrained, and the impact of the generated data on model performance was systematically evaluated. The recognition rates increased by 1.69% and 3.03%, respectively, thereby validating the effectiveness of the generated data

Topics & Concepts