Professor Thomas Huang and his research group have developed a new method of recovering high-resolution images from their low-resolution observations. They developed a Recurrent Neural Network Model for image super-resolution which fully exploits sparse representations prior while keeping all the benefits of deep networks.
They trained a cascade of RNNs, each with a fixed upscaling factor, enabling high resolution images upscaled by large ratios to be obtained with less training efforts and better recovery quality This technique had higher Peak Signal to Noise ratio than the best state of the art methods by up to 1dB.