Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Kernel Adversarial Learning for Real-world Image Super-resolution

Hu Wang, Congbo Ma, Jianpeng Zhang, Wei Emma Zhang, Gustavo Carneiro

TL;DR

A more realistic process to synthesise low-resolution images for real-world image SR by introducing a new Kernel Adversarial Learning Super-resolution (KASR) framework, where degradation kernels and noises are adaptively modelled rather than explicitly specified.

Abstract

Current deep image super-resolution (SR) approaches aim to restore high-resolution images from down-sampled images or by assuming degradation from simple Gaussian kernels and additive noises. However, these techniques only assume crude approximations of the real-world image degradation process, which should involve complex kernels and noise patterns that are difficult to model using simple assumptions. In this paper, we propose a more realistic process to synthesise low-resolution images for real-world image SR by introducing a new Kernel Adversarial Learning Super-resolution (KASR) framework. In the proposed framework, degradation kernels and noises are adaptively modelled rather than explicitly specified. Moreover, we also propose a high-frequency selective objective and an iterative supervision process to further boost the model SR reconstruction accuracy. Extensive experiments validate the effectiveness of the proposed framework on real-world datasets.

Kernel Adversarial Learning for Real-world Image Super-resolution

TL;DR

A more realistic process to synthesise low-resolution images for real-world image SR by introducing a new Kernel Adversarial Learning Super-resolution (KASR) framework, where degradation kernels and noises are adaptively modelled rather than explicitly specified.

Abstract

Current deep image super-resolution (SR) approaches aim to restore high-resolution images from down-sampled images or by assuming degradation from simple Gaussian kernels and additive noises. However, these techniques only assume crude approximations of the real-world image degradation process, which should involve complex kernels and noise patterns that are difficult to model using simple assumptions. In this paper, we propose a more realistic process to synthesise low-resolution images for real-world image SR by introducing a new Kernel Adversarial Learning Super-resolution (KASR) framework. In the proposed framework, degradation kernels and noises are adaptively modelled rather than explicitly specified. Moreover, we also propose a high-frequency selective objective and an iterative supervision process to further boost the model SR reconstruction accuracy. Extensive experiments validate the effectiveness of the proposed framework on real-world datasets.

Paper Structure

This paper contains 14 sections, 7 equations, 5 figures, 10 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Real-world Image Super-resolution
  4. Adversarial Training
  5. The Proposed Method
  6. Overall Framework
  7. Kernel Adversarial Noise Simulation
  8. Objective Functions
  9. Experiments
  10. Experimental Setup
  11. Overall Performance
  12. Ablation Study
  13. Analyses
  14. Conclusion

Figures (5)

  • Figure 1: Intuitive visualisation of the original LR image and degraded LR image produced by the KASR model. The image in (c) highlights (in pink) the differences between the original low-resolution image in (a) and the LR image down-sampled by the proposed KASR model in (b). Notice that the differences are mainly due to colour changes combined with some texture distortions.
  • Figure 2: The overall framework of the proposed KASR framework. The framework consists of three parts: (a) the Kernel Adversarial Noise Simulation (KANS) to adaptively simulate the image degradation process; (b) the High-frequency Selective Objective to force the model to focus on high-frequency regions within images due to the higher importance placed in the image regions containing high-frequency for the SR task; and (c) the stacking of multiple KANS modules, where the Iterative Supervision (IS) can leverage the supervision signals to accurately refine the SR image reconstruction.
  • Figure 3: Illustration of the differences between original images $I_{LR}$ and KASR produced images $I'_{LR}$.
  • Figure 4: Visual comparison between the original HR image, Bicubic interpolation, RCAN and our proposed KASR model.
  • Figure 5: Visual comparison between the original HR image, Bicubic interpolation, Noise-injection and the proposed model.