Single-frame image super-resolution processing method based on diffusion model

Abstract

The invention discloses a single-frame image super-resolution processing method based on a diffusion model. The method comprises the following steps: establishing a diffusion process and a conditional noise predictor of a super-resolution diffusion model; inputting a high-low resolution image data set into a conditional noise predictor of the super-resolution diffusion model, and training the conditional noise predictor by using the diffusion process of the super-resolution diffusion model to obtain a pre-trained conditional noise predictor; inputting a to-be-processed low-resolution image into an inverse process of the super-resolution diffusion model to obtain a residual prediction image; and adding the residual prediction image and the low-resolution image subjected to the up-sampling operation to obtain a predicted super-resolution image. According to the invention, multiple super-resolution images with consistent features are generated for the same low-resolution image, each super-resolution image has rich textures and naturalness, the over-smoothing problem and the mode collapse problem are avoided, and the reconstruction performance is improved with less time overhead.

Description

technical field [0001] The invention belongs to an image processing method in the fields of image processing technology, computer vision, and deep learning, and in particular relates to a diffusion model-based single-frame image super-resolution processing method in the field of image super-resolution technology. Background technique [0002] Single-frame image super-resolution tasks have a wide range of applications in the field of computer vision, such as target recognition, multimedia technology, aerospace imaging, etc. The single-frame image super-resolution task aims to recover high-resolution images from low-resolution images, which is a pathological problem because multiple high-resolution images can degenerate into the same low-resolution image. [0003] In order to establish the mapping between high-resolution images and low-resolution images, many neural network-based methods have emerged in recent years, and these methods can be divided into three main types: meth...

AI Technical Summary

Problems solved by technology

However, the PSNR-oriented feedforward super-resolution neural network method obtains the average of multiple super-resolution potential solutions, resulting in a super-resolution image that is too smooth; the method based on GAN is prone to "mode collapse", resulting in the generation of super-resolution images. Close to the same solution; flow-based methods require huge model parameters and computational overhead

How to generate multiple high-quality high-resolution images consistent with the same low-resolution image while avoiding large computational overhead is not solved and realized in the prior art

Images