An image super-resolution reconstruction method based on quality evaluation and feature statistics

Abstract

According to the invention, a super-resolution reconstructed picture is more consistent with a human eye visual perception effect; at the same time, the feature structure of the picture is maintained.The invention provides an image super-resolution reconstruction method based on no-reference quality evaluation and feature statistics. An adversarial learning network model is constructed, more high-frequency detail edges are generated, the image looks like more textures, the tone, the brightness and the sharpness of the image are adjusted through no-reference quality evaluation, so that the generated image is closer to human visual perception, and the internal feature structure of the image is maintained through feature statistics. Compared with a traditional super-resolution reconstructionmethod, the high-resolution image generated through the method has more abundant real texture details, the human eye perception effect of the image is improved, and the result content of the image cannot be damaged.

Description

technical field [0001] The invention relates to the fields of computer vision, image super-resolution reconstruction, and image quality evaluation, in particular to an image super-resolution reconstruction method based on no-reference quality evaluation and feature statistics. Background technique [0002] Image super-resolution reconstruction is to restore high-resolution images from low-resolution images with limited information. However, super-resolution reconstruction is a one-to-many pathological problem, because multiple high-resolution images may correspond to the same low-resolution image. Thanks to the development of the current popular deep learning, this type of method limits the solution space by learning the corresponding mapping relationship from the super-resolution dataset, thereby alleviating this ill-conditioned problem to a certain extent, and the super-resolution reconstruction effect A major breakthrough has been made. Most of these works unilaterally ...

AI Technical Summary

Problems solved by technology

This method overcomes the common overfitting and underfitting problems in super-resolution reconstruction to a certain extent, but the GAN network training is more complicated, and the number of parameters is large, which can easily lead to some false edges in the generated image and affect the realness of the image. look and feel

[0007] Although the perceptual loss and the confrontation loss are applied to image super-resolution reconstruction, so that the generated image has a certain perceptual effect, but the generated result faces too many false edges, structural damage, etc.

Images