Variable-magnification image super-resolution network model

Abstract

The invention discloses a variable-magnification image super-resolution network model. The model comprises a parameterized residual error learning network PRNet, a residual error refined learning network RRNet and a superposition network. The parameterized residual error learning network PRNet is used for learning mapping from a low-resolution LR image to ahigh-resolution HR image; the residual error refined learning network RRNet is used for learning the mapping from a reconstructed high-resolution image to the residual error image; and the superposition network is used for superposing the high-resolution HR image and the residual image to form a final super-resolution SR image and outputting the super-resolution SR image. Explicit expression is carried out on the magnification parameters, and the parameterized residual learning network model is established, so that the model can accept input of any scale, and the variable magnification super-resolution task requirement is met. A residual refinement learning network is provided, and the mapping relations between the reconstructed high-resolution image and the reconstructed residualare further learned, so that residual compensation is carried out on the reconstructed image, and the super-resolution reconstruction quality is improved.

Description

technical field [0001] The invention belongs to the technical field of digital image processing and relates to an image super-resolution network model, in particular to an image super-resolution network model with variable magnification. [0002] technical background [0003] Deep learning technology has promoted a huge leap in image super-resolution (SR) reconstruction performance, but the deep learning model for super-resolution reconstruction can only target a fixed magnification at a time. If different magnifications are to be performed simultaneously For super-resolution tasks, it is necessary to train multiple deep learning network models corresponding to different magnifications. This limitation of super-resolution networks restricts their practical applications for applications that need to perform arbitrary indeterminate magnifications. [0004] Super-resolution application scenarios with uncertain magnification are ubiquitous in reality. A common scenario is face s...

AI Technical Summary

Problems solved by technology

[0006] It can be seen that for the super-resolution task of variable magnification, the current two processing methods have serious defects, and the way of training the corresponding network model for each multiple is cumbersome and impractical; first interpolate to a uniform resolution and then super-resolution reconstruction The method ignores the differences in content components of input images with different resolutions, and it is difficult to obtain ideal reconstruction effects.

The reason why a network trained for one magnification (such as X2) cannot be applied to another magnification (such as X3) is that the existing super-resolution reconstruction network does not explicitly express the variable of magnification. Therefore, it is necessary to propose a super-resolution network model with variable magnification to meet the requirements of arbitrary scale input

Images