ScholarWorks@중앙대학교

초록

In this study, we propose a deep learning-based superresolution network for reconstructing high-resolution synthetic aperture radar (SAR) images under bandwidth-limited conditions. In general, automotive SAR systems operate under strict bandwidth regulations, which impose a limitation on enhancing range resolution. To address this issue, we design a generative adversarial network (GAN)-based super-resolution method that enables high-resolution image generation without hardware modifications. The proposed network adopts a GAN architecture consisting of a generator and a discriminator, and is trained to generalize across diverse environments using data collected with a TI AWR1642 radar. The training optimizes a combination of various losses to promote both structural fidelity and perceptual quality in generated SAR images. Through this approach, the proposed model achieves notable performance improvements. In particular, compared to the bicubic interpolation method, the proposed model increases the peak signal-to-noise ratio by 20.86 dB, improves the structural similarity index by 0.44, and reduces the learned perceptual image patch similarity by 0.48. Moreover, in real-time autonomous driving scenarios, it maintains competitive performance against other GAN-variant model. In addition, the proposed super-resolution method reduces the half-power bandwidth by 82.39%, that reduction is 50.01%p greater than that achieved by the Unet baseline.

키워드