Methods for deep-learning based super-resolution using high-frequency loss

A machine learning model can be trained to perform super-resolution by using high-frequency loss. One or more degradations of a first type can be applied to reference images to generate corresponding degraded images that include a reduced amount of high-frequency texture information when compared to the corresponding reference images. A mapping function associated with a machine learning process can used to generate predicted images. One or more degradations of a second type can be applied to the predicted images and the reference images to generate corresponding low-frequency images. The low frequency images corresponding to the predicted images can be compared to the low-frequency images corresponding to the reference images. Based at least partially on the comparison, a loss value can be calculated. If the loss value exceeds a loss value threshold, the mapping function can be updated in accordance with the loss value.