Serial GANs: A Feature-Preserving Heterogeneous Remote Sensing Image Transformation Model

In recent years, the interpretation of SAR images has been significantly improved with development deep learning technology, and using conditional generative adversarial nets (CGANs) for SAR-to-optical transformation, also known as image translation, become popular. Most existing translation methods based on are modified CycleGAN pix2pix, focusing style transformation in practice. addition, optical characterized by heterogeneous features large spectral differences, leading to problems such incomplete details distortion urban or semiurban areas complex terrain. Aiming solve Serial GANs, a feature-preserving remote sensing model, is proposed this paper first time. This model uses Despeckling GAN Colorization complete transformation. transforms into gray images, retaining texture semantic information. obtained step color keeps structural unchanged. The provides new idea Through decoupling network design, detail information relatively independent process thereby enhancing generated reducing its distortion. Using SEN-2 satellite reference, compares degree similarity between different models results revealed that obvious advantages feature reconstruction economical volume parameters. It showed GANs have great potential