Optical Image Guided Multi-scale Learning for Synthetic Aperture Radar Image Super-resolution (preprint)

ABSTRACT

Synthetic Aperture Radar (SAR) image Super-Resolution (SR) has various application including military reconnaissance and marine monitoring. Comparing with other natural images, the SAR images suffer from lower spatial resolution and less amplitude information, which provide insufficient information for SAR image reconstruction. Therefore, current SAR image SR methods show worse performance. In this paper, considering optical images have higher resolution and richer texture information, we propose a multi-scale learning based optical image guidance network (MLOG) to make good use of high-resolution (HR) optical image for SAR image SR. Specifically, we design an optical image weighted guidance network (OWG) to get more discriminative feature and reduce the impact of the lack of detailed information in LR SAR images. Then a multi-scale learning network (MLN) is proposed to obtain the global features from optical image. After that, a residual multi-scale block (RMB) is introduced to extract the global multi-scale context information. Finally, the extracted multi-scale features are reconstructed to generate high-resolution SAR images through a convolutional reconstruction layer. The experimental results show that the proposed MLOG method achieves state-of-the-art performance for SAR image SR on public available datasets.