Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
175 tokens/sec
GPT-4o
7 tokens/sec
Gemini 2.5 Pro Pro
42 tokens/sec
o3 Pro
4 tokens/sec
GPT-4.1 Pro
38 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

Unpaired Image Super-Resolution using Pseudo-Supervision (2002.11397v1)

Published 26 Feb 2020 in eess.IV and cs.CV

Abstract: In most studies on learning-based image super-resolution (SR), the paired training dataset is created by downscaling high-resolution (HR) images with a predetermined operation (e.g., bicubic). However, these methods fail to super-resolve real-world low-resolution (LR) images, for which the degradation process is much more complicated and unknown. In this paper, we propose an unpaired SR method using a generative adversarial network that does not require a paired/aligned training dataset. Our network consists of an unpaired kernel/noise correction network and a pseudo-paired SR network. The correction network removes noise and adjusts the kernel of the inputted LR image; then, the corrected clean LR image is upscaled by the SR network. In the training phase, the correction network also produces a pseudo-clean LR image from the inputted HR image, and then a mapping from the pseudo-clean LR image to the inputted HR image is learned by the SR network in a paired manner. Because our SR network is independent of the correction network, well-studied existing network architectures and pixel-wise loss functions can be integrated with the proposed framework. Experiments on diverse datasets show that the proposed method is superior to existing solutions to the unpaired SR problem.

Citations (146)

Summary

  • The paper introduces a novel unpaired image super-resolution technique using a dual-network GAN framework with an unpaired kernel/noise correction network and a pseudo-paired SR network.
  • Experimental results show superior performance over state-of-the-art unpaired methods on various synthetic and real-world datasets, improving quantitative and qualitative metrics.
  • This approach advances practical unpaired SR by enabling training without paired HR-LR data, expanding feasibility for real-world deployment where capturing pairs is difficult.

Unpaired Image Super-Resolution Using Pseudo-Supervision

Unpaired image super-resolution poses significant challenges in the field of low-level vision, mainly due to the absence of paired high-resolution (HR) and low-resolution (LR) image datasets. Traditional methods often rely on generating paired datasets through deterministic downscaling operations, which fail to capture the complex degradation processes observed in real-world LR images. The paper introduces a novel unpaired super-resolution (SR) technique leveraging generative adversarial networks (GANs), specifically designed to operate without the necessity of paired/aligned training datasets.

Methodology

The proposed approach introduces a dual-network framework comprising an unpaired kernel/noise correction network and a pseudo-paired SR network. The former, based on CycleGAN architectures, is responsible for translating real-world LR images into a clean LR domain, correcting noise and kernel discrepancies without assuming specific degradation forms. Meanwhile, the latter employs a mapping from these corrected pseudo-clean LR images to HR images, utilizing paired training derived from available HR images downscaled by predetermined methods, effectively using pseudo-supervision.

Central to this approach is the separation between the noise correction network and the SR network. This modularity enables the integration of any existing state-of-the-art SR architectures and pixel-wise loss functions, providing flexibility and compatibility with established methods.

Experimental Results

Extensive experiments were executed across diverse datasets to validate the efficacy of the proposed method. On synthetically degraded and real-world datasets, the results demonstrated superior performance in comparison to other state-of-the-art unpaired SR techniques. In particular, evaluations on the DIV2K realistic-wild dataset and large-scale facial and aerial image datasets showed remarkable improvements in both quantitative metrics, such as PSNR and SSIM, and qualitative assessments, especially when the perceptual quality of the images was considered.

Implications and Future Directions

This research offers a method of advancing unpaired SR in practical, real-world applications, bridging a crucial gap between lab-based models and field utility. By allowing SR models to train on datasets without predefined HR-LR pairings, this approach significantly enhances the feasibility of deploying SR technology in environments where capturing paired datasets would be impractical or impossible.

Future work could explore the robustness of this framework across varying degradation types and intensities, as well as automate hyperparameter tuning to optimize performance. Additionally, enhanced perceptual losses and integration with more advanced GAN architectures could further improve image quality and training stability.

Concluding Remarks

This paper presents a forward-thinking approach in image super-resolution, emphasizing flexibility and adaptability through pseudo-supervised learning. It provides a pathway for richer, more practical SR applications and sets a new benchmark for unpaired SR methodologies. The integration of this framework into real-world systems could improve various applications in fields ranging from medical imaging to satellite photo enhancement, demonstrating its broad potential impact.

X Twitter Logo Streamline Icon: https://streamlinehq.com