Image super-resolution enhancement method based on knowledge distillation

Abstract

The invention discloses an image super-resolution enhancement method based on knowledge distillation. The method comprises the following steps of 1) obtaining training data and test data; 2) traininga teacher network, wherein the teacher network is provided with a neural network model of a relatively deep convolutional layer; 3) training a student network; 4) performing guidance learning on the student network by the teacher network, and enabling the student network to learn and absorb feature graphs of the teacher network through three sets of guidance experiments; 5) testing and evaluatingan image reconstruction effect; and 6) further guiding the student network according to different matrix relations between the output feature graphs. By means of related thoughts of the knowledge distillation, the performance of the teacher network is transmitted to the student network; the student network model can efficiently run in mobile equipment and embedded equipment with low-power-consumption limitation; and on the premise that the network structure of the student network is unchanged, the PSNR of the student network guided by the teacher network is obviously improved, and a better reconstruction effect is obtained.

Description

technical field [0001] The invention relates to the fields of computer vision and deep learning, in particular to an image super-resolution enhancement method based on knowledge distillation. Background technique [0002] Super-resolution (Super Resolution, SR) is a classic problem in computer vision. The purpose of single image super-resolution (Single Image Super-Resolution, SISR) is to use methods such as digital image processing, from a single low-resolution (Low- The corresponding high-resolution (High-Resolution, HR) image is recovered from the Resolution, LR) image. In the super-resolution problem, assuming a low-resolution image X, our goal is to recover a super-resolution image Y′ that is as similar as possible to the ground truth (Ground Truth, GT) image Y. [0003] Traditional interpolation-based magnification methods, including Bilinear Interpolation and Bicubic Interpolation, use fixed calculation formulas to perform weighted averages using neighborhood pixel i...

AI Technical Summary

Problems solved by technology

Although increasing the depth of the network can bring better super-resolution reconstruction results, but at the same time the amount of computation and memory consumption will also increase. In many practical application scenarios (such as mobile terminals and embedded low-power consumption constraints ), unable to run deep convolutional neural network models in real time

Images