Lightweight image super-resolution reconstruction method based on multi-dimensional knowledge distillation

Abstract

The invention discloses a lightweight image super-resolution reconstruction method based on multi-dimensional knowledge distillation. The method comprises the following steps: preprocessing a teacher network; preprocessing the data set to generate a low-resolution image; inputting the low-resolution image into the constructed student network, and outputting a super-resolution image reconstructed by the student network; inputting the low-resolution image into a plurality of teacher networks to obtain a reconstructed super-resolution image group; respectively carrying out L1 loss calculation and perception loss calculation on the super-resolution image group of student network reconstruction and the super-resolution image group of teacher network reconstruction, and carrying out update training of back transmission to obtain a final student network model; and inputting the low-resolution image into the final network model, and outputting a super-resolution image. According to the super-resolution training mode, the parameter quantity is reduced, indexes and visual effects which are comparable to those of a fully-supervised teacher network trained by paired data are obtained, meanwhile, the size of the model is effectively reduced, and compared with a traditional training mode, the model is obviously improved.

Description

technical field [0001] The invention relates to the technical field of image super-resolution reconstruction, in particular to a lightweight image super-resolution reconstruction method based on multi-dimensional knowledge distillation. Background technique [0002] As a second-generation image restoration technology, super-resolution reconstruction mainly converts low-resolution images into high-resolution clear images. Learning-based super-resolution methods are mainly divided into two categories: fidelity-oriented methods for reconstructed images and perceptual quality-based methods. The former is guided by the quality evaluation between the super-resolution reconstructed image and the original image, and the goal is to produce high result objective indicators such as peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) to ensure that it is higher than the original image. similarity, but the visual perception is usually poor, mainly manifested as too ...

AI Technical Summary

Problems solved by technology

This downsampling method does not fully reflect the real image degradation, so the neural network based on this training often performs poorly on real-world image reconstruction tasks

At the same time, the traditional training method adopts the constraint method of single-dimensional multi-loss function, that is, only paired data sets are used, resulting in too large solution space from high-resolution images to low-resolution images, ignoring the multi-dimensional data sets for The role of constrained solution spaces, network learning difficulties

In order to enrich the detailed texture, it is often through deepening the network, adding dense connections, etc., but with a good visual effect, the amount of calculation is huge, which consumes a lot of resources.

Images