Generative Adversarial Super-Resolution at the edge with knowledge distillation

Single-Image Super-Resolution can support robotic tasks in environments where a reliable visual stream is required to monitor the mission, handle teleoperation or study relevant details. In this work, we propose an efficient Generative Adversarial Network model for real-time Super-Resolution, called EdgeSRGAN1. We adopt tailored architecture of original SRGAN and quantization boost execution on CPU Edge TPU devices, achieving up 200 fps inference. further optimize our by distilling its knowledge smaller version network obtain remarkable improvements compared standard training approach. Our experiments show that fast lightweight preserves considerably satisfying image quality heavier state-of-the-art models. Finally, conduct transmission with bandwidth degradation highlight advantages proposed system mobile applications.