SR-DNnet: A Deep Network for Super-Resolution and De-Noising of ISAR Images

Abstract

Inverse synthetic aperture radar (ISAR) images have become one of the most important pieces of information for airborne and maritime target identification. In general, ISAR images with higher resolution and lower background noise provide more precise target information, thus improving target identification accuracy. However, upgrading the resolution of the ISAR system is costly. Super-resolution algorithms that can utilize low-resolution echoes to obtain high-resolution imaging results have become an important means of improving ISAR imaging resolution. The traditional ISAR super-resolution imaging technique suffers from high side lobes and wide main lobes. In addition, denoising algorithms based on filtering operators tend to lead to image blurring. This work proposes a deep network for super-resolution and de-noising of ISAR images called SR-DNnet. Specifically, we view super-resolution and de-noising as a series of up-sampling, two-dimensional filtering, and threshold shrinkage. These operations are exactly what deep networks are good at. SR-DNnet has 15 layers, enabling 4x super-resolution and de-noising of ISAR images. The parameter scale of SR-DNnet is much smaller than most deep networks, which makes it efficient to train. The SR-DNnet we built features complex-value inputs, residual learning, multipath learning, and progressive up-sampling. A series of simulated and measured dataset experiments prove that the SR-DNnet is efficient and well-performed on super-resolution and de-noising.