Evaluation of Simulated SAR Images Created by Generative Models Based on Pixel-wise Similarity

In recent years, deep learning has become a research hotspot when processing Synthetic Aperture Radar(SAR) data. Due to high cost of SAR image acquisition, many generative models are proposed to simulate SAR images. However, it's still hard to assess which algorithm performs better. We conduct a multi-faced large-scale empirical study on generative models including Generative Adversarial Nets(GANs) and Encoder-Decoder based models for simulating SAR images and evaluate performance of these models using pixel-wise similarity. Results show that the difference is small in contour similarity for generated SAR images produced by various of networks, however, there exists sharp difference in terms of pixel correlation and Fréchet Inception Distance(FID). The Variant Autoencoder network has certain advantages in many indicators under limited training data. Pixel correlation is rather accurate taking into account multi-features of SAR images in low level dimension and provides meaningful guidance in evaluating quality of SAR images.