Height Estimation from a Single SAR Image by Depth-Aware Proxyless Neural Architecture Search

Abstract

Height estimation can be estimated from a single synthetic aperture radar (SAR) Image via an advanced U-shaped convolution neural network (CNN)—Unet. Manually designed architectures are applied to achieve high reconstruction accuracy. Neural Architecture Search (NAS) is efficient for designing the architecture of convolution neural network (CNN) automatically for a specific task, but it is difficult to be deployed directly due to the limitations of computational resources. Then, proxy tasks are developed for traditional NAS methods. However, the trained model for proxy tasks is not ptimal for the target task. In this paper, we propose a depth-aware penalty proxyless NAS method for Unet (DPNAS-Unet) to estimate the height from a single SAR Image. Symmetric residual blocks are proposed to improve the representation capability of the model. A depth-aware architecture loss is proposed for optimal architecture learning. Supplementary sparse height map is included as inputs of our model to improve the performance. In the end, comparison experiments are performed to show the superiority of PDNAS-Unet.