Transitive Learning: Exploring the Transitivity of Degradations for Blind Super-Resolution

Yuanfei Huang, Jie Li, Yanting Hu, Xinbo Gao, Wen Lu

Introduction

Single Image Super-Resolution (SISR) aims to reconstruct high-resolution (HR) images from low-resolution (LR) inputs. While deep learning models perform exceptionally well when the degradation process (e.g., bicubic downsampling) is known, real-world "blind" super-resolution remains a challenge due to unknown blur kernels and noise levels. Existing approaches often rely on high-dimensional degradation representations or slow, iterative "zero-shot" optimization.

This paper introduces Transitional Learning for blind Super-Resolution (TLSR). The core idea is to treat unknown degradations as intermediate "transition states" between known primary degradations. By estimating a 1D scalar called the Degree of Transitionality (DoT), the model adaptively reconfigures its internal weights via network interpolation, achieving state-of-the-art accuracy with significantly higher inference speeds.

The Concept of Transitionality

The authors propose that many common degradations, such as Gaussian blur or additive white Gaussian noise (AWGN), can be modeled as points on a 1D manifold between two "primary" states (e.g., no blur vs. maximum blur). This reduces the complex task of degradation estimation to finding a single scalar value, $\tau \in [0, 1]$.

Visualizing transitional degradations for kernels and noise.

As shown in the figure above, synthesized kernels ($\mathcal{B}^\tau$) and noises ($\mathcal{N}^\tau$) can effectively approximate ground truth degradations by interpolating between primary states. This 1D representation simplifies the search space and enhances model interpretability.

The TLSR Framework

The TLSR architecture consists of three specialized components designed to handle these transitional states efficiently:

  1. DoT Estimation Network (DoTNet): A lightweight module that processes random crops of the LR image to predict the scalar $\hat{\tau}$. Using crops makes the estimation step extremely fast and robust to local variations.
  2. Homogeneous Feature Extraction: A shared backbone (using Residual Blocks) that extracts features from the LR image regardless of the specific degradation level.
  3. Transitional Learning Module: This module implements Network Interpolation. Instead of fixed weights, it uses $\hat{\tau}$ to linearly blend the weights of two sets of filters trained on the primary degradations. This allows the network to dynamically adapt its behavior to the specific degradation level of the input image.
The complete TLSR architecture, including DoT estimation, feature extraction, and the transitional learning module.

Key Insights and Results

The transitional learning approach offers several advantages:

  • Efficiency: By interpolating weights rather than features, the model avoids the computational overhead of multi-branch architectures. TLSR is approximately 10x faster than IKC and 2000x faster than ZSSR for $\times2$ upscaling.
  • Accuracy: TLSR demonstrates superior performance on benchmark datasets. For instance, on Set14 with blur degradations, it outperformed IKC by 0.69dB in PSNR.
  • Interpretability: The scalar $\tau$ provides a direct measure of the estimated degradation severity, offering insights into the model's decision-making process.

Conclusion

TLSR bridges the gap between efficient integrated learning and accurate but slow zero-shot methods. By leveraging the transitivity of degradations, it provides a scalable and interpretable framework for blind super-resolution. Future work may extend this 1D transition logic to multi-axis transitions to handle complex, mixed degradations in a single-stage framework.

Summary updated

BibTeX
@article{huang2023transitive,
  author = {Yuanfei Huang and Jie Li and Yanting Hu and Xinbo Gao and Wen Lu},
  title = {{Transitive Learning: Exploring the Transitivity of Degradations for Blind Super-Resolution}},
  journal = {IEEE Trans. Pattern Anal. Mach. Intell.},
  year = {2023},
  doi = {10.1109/tpami.2022.3206870},
}
Figure