| Title |
Meta-learning with Pixel-wise Adaptive Weight for Robust Image Super-resolution |
| Authors |
장종환(Jong-Hwan Jang) ; 최장훈(Janghoon Cho) |
| DOI |
https://doi.org/10.5573/ieie.2024.61.3.93 |
| Keywords |
Deep learning; Meta learning; Image super resolution; Adaptive weight map |
| Abstract |
Image super-resolution(SR) is a computer vision task that converts low-resolution images into high-resolution images. With the advent of deep learning, many new and effective methods have been proposed. However, most SR methods are conducted under the assumption that the low-resolution degradation process is bicubic downsampling, so there is a problem that they are not well applied to real images with complex degradation types. In this work, we propose a method to improve the performance of Single Image Super-Resolution that is robust against various degradation types, without changing the architecture of conventional SR network, by utilizing the internal information. For utilizing the internal information, we adopt a meta-learning algorithms using a meta-learner network, which constructs an pixel-wise adaptive weight map(PAW) tailored to the given input image. This approach can be quickly applied to bicubic downsampling kernels as well as blind downsampling kernels and improves performance by encouraging the network to focus more on complex parts of the image. Experiments on various benchmark SR datasets show that our proposed method improves the performance while maintaining the structure of the existing SR network. |