Most current super-resolution methods rely on a pair of low and high-resolution images to train a network in a supervised manner. However, in real-world scenarios, such pairs are not available. Instead of directly addressing this problem, most tasks employ the popular bicubic down-sampling strategy to generate low-resolution images artificially. Unfortunately, this strategy introduces more artifacts, removing natural incense and other real-world characteristics. Moreover, super-resolution networks trained on such bicubic images suffer many struggles to generalize the natural images.

Read more: https://analyticsindiamag.com/guide-to-image-super-resolution-by-esrgan/

#artificial-intelligence #machine-learning