Co-learning method of super-resolution network based on clique structure subbands in wavelet domain

Abstract

The invention discloses a method for learning super-resolution network based on sub-band of group structure in wavelet domain, which realizes super-resolution restoration of single picture by constructing a novel super-resolution network comprising a feature extraction network and an image restoration network. The input low-resolution image is represented as a set of feature maps by a feature extraction network. The image restoration network includes a cluster upsampling module and a convolution layer. The output features of the feature extraction network are transferred to the cluster upsampling module for sub-band extraction, self-residual learning and sub-band refinement. The inverse discrete wavelet transform is used to change the resolution of the feature image so as to recover the feature image with high resolution. Then the high resolution image is restored by convolution layer. The method of the invention has good quantization index and can effectively realize super-resolutionrestoration of a single picture.

Description

technical field [0001] The invention relates to computer vision image super-resolution restoration technology, in particular to a method for jointly learning a super-resolution network based on sub-bands of a clique structure in a wavelet domain, which can solve the problem of super-resolution restoration of a single picture. Background technique [0002] Single image super-resolution refers to reconstructing a high-resolution image from a single low-resolution image, which is an ill-posed inverse problem. For several years, there has been increasing research interest in single image super-resolution. Recently, convolutional neural networks (refs [1], [2], [3]) have significantly improved the peak signal-to-noise ratio of reconstructed high-resolution maps and real images in the single-image super-resolution problem. These networks usually first use a feature extraction module to extract a series of feature maps from a low-resolution image, and then cascade with an upsampli...

AI Technical Summary

Problems solved by technology

[0008] However, these networks often tend to produce blurred and over-smoothed high-resolution images, often lacking texture details, and it is difficult to restore the details of a single high-definition image, resulting in low peak signal-to-noise ratio and low image similarity structure index

Images