An image super-resolution restoration method, system and device

Abstract

This application discloses a method, system, and device for super-resolution image restoration. It obtains a discrete reference distribution of a high-resolution image, then trains a U-net network in a diffusion model using different low-resolution images to obtain the predicted category distribution corresponding to the high-resolution predicted label map. The prediction bias of the U-net network is obtained through the discrete reference distribution, the predicted category distribution, and multiple preset functions. If the bias is large, low-resolution images in the training set are added and the network is trained again until the bias is no greater than a preset threshold. This allows the U-net network to better adapt to the multi-modal characteristics of the real posterior distribution, avoiding problems caused by the mismatch in the U-net network's expressive power. The low-resolution image to be restored is input into the trained U-net network to obtain a high-resolution label map, which is then transcoded to obtain a high-resolution image, effectively improving the quality of the restored image and reducing structural degradation and artifacts.

Images