BluRef: Unsupervised Image Deblurring with Dense-Matching References

1. Problem Statement

Image deblurring is critical for restoring visual quality and enabling downstream computer vision tasks. However, current state-of-the-art approaches face significant limitations:

• Supervised Methods: Require paired datasets (blurry image + corresponding sharp ground truth). Acquiring such data is expensive and logistically difficult, often requiring specialized hardware (beam splitters, synchronized cameras) that is impractical for real-world scenarios like dashcams or bodycams.
• Unsupervised/Reblurring Methods: Existing unsupervised approaches often rely on "reblurring" strategies (mapping blur to an intermediate domain) or synthetic data augmentation. These suffer from domain gaps, where the synthetic or intermediate blur does not match the target real-world blur, leading to poor generalization and performance.

Core Question: Can high-performing deblurring models be trained entirely within the target domain using only unpaired data, eliminating the need for paired ground truth or complex multi-stage pipelines?

2. Methodology: BluRef

The authors propose BluRef, a novel unsupervised framework that learns from unpaired blurry images and sharp reference images (from similar scenes) to generate pseudo-ground truth. The method operates in an iterative refinement loop.

A. Core Components

1. Unpaired Reference Images: Instead of paired data, BluRef utilizes a set of $N$ sharp reference images $\{I_{ref}^n\}$ from the same scene distribution as the blurry input $I_{blur}$. These references do not need to be geometrically aligned or captured simultaneously.
2. Dense Matching (DM) Model: A pre-trained model (based on PDC-Net+ and GLU-Net) trained via self-supervision on synthetically warped sharp images. This model establishes pixel-level correspondences between the deblurred estimate and the sharp references, even across significant geometric and appearance variations.
3. Iterative Training Loop:
   • Step 1: Pseudo-Sharp Generation: The current deblurred estimate $I_{deblur}^{(k)}$ is matched against the sharp references using the DM model. This generates a pseudo-sharp image ($I_{pseudo}^{(k)}$) and a confidence mask ($M_{pseudo}^{(k)}$). The mask suppresses unreliable correspondences.
   • Step 2: Network Optimization: The deblurring network $D$ is trained to minimize the reconstruction loss between its output and the generated pseudo-sharp target, weighted by the confidence mask.
   • Refinement: The process repeats. As the deblurring network improves, the pseudo-sharp targets become more accurate, creating a positive feedback loop.

B. Aggregation Strategies

When multiple reference images are available ($N > 1$), the paper proposes three strategies to aggregate matches:

1. Weighted Average: Independent matching for each reference, then averaging results.
2. Sequential Accumulation: Iteratively refining the pseudo-image by using the previous iteration's output as input for the next match.
3. Progressive Reference Averaging (Best Performance): Combines sharp details from previous iterations while selectively enhancing unmatched areas with new references. This ensures continuity of detail without discarding previously learned sharpness.

C. Training and Inference

• Training: Requires only unpaired video frames. The DM model is pre-trained on synthetic warps (no paired blur/sharp data needed for the main deblurring task).
• Inference: Once trained, the dense matching module and reference images are discarded. The model acts as a standard, single-pass deblurring network, maintaining the computational efficiency of supervised models.

3. Key Contributions

1. First Unpaired Reference-Guided Framework: BluRef is the first method to use reference images to generate pseudo-ground truth for unsupervised deblurring, eliminating the need for paired datasets or pre-trained deblurring models.
2. Scalable Pseudo-Data Generation: The iterative process generates high-quality pseudo-pairs that can be reused to train models of varying capacities (including lightweight models for mobile devices), extending utility beyond the specific backbone used during training.
3. State-of-the-Art Performance: The approach achieves performance comparable to, and in some cases exceeding, supervised models, despite being trained without ground truth supervision.

4. Experimental Results

The authors evaluated BluRef on GoPro (synthetic blur), RB2V (real-world street blur), and PhoneCraft (unpaired smartphone video).

• Quantitative Performance:
  • On GoPro ($\Delta=1$ frame), BluRef (NAFNet backbone + Progressive Averaging) achieved 31.94 dB PSNR, approaching the supervised upper bound of 33.32 dB.
  • On RB2V, BluRef (Restormer backbone) achieved 27.87 dB PSNR, exceeding the supervised Restormer upper bound (27.43 dB).
  • It significantly outperformed existing unsupervised baselines (DualGAN, UID-GAN, UAUD) by a large margin.
• Robustness: The method remains stable even when the temporal distance ($\Delta$) between the blurry input and reference frames increases (up to 20 frames), demonstrating resilience to significant motion and content changes.
• Qualitative Results: Visual comparisons show BluRef successfully recovers fine details (e.g., facial features, tree branches) and body shapes that other unsupervised methods fail to restore, producing results visually close to the ground truth.

5. Significance

• Democratizing Deblurring: By removing the dependency on costly paired data collection, BluRef makes high-quality deblurring accessible for real-world applications where paired data is impossible to acquire (e.g., autonomous driving, surveillance).
• Efficiency: Unlike reblurring-based methods that require complex multi-stage pipelines, BluRef offers a direct inference path, making it suitable for deployment on resource-constrained devices.
• Paradigm Shift: It demonstrates that "domain gap" issues can be mitigated by leveraging unpaired references within the target domain itself, rather than forcing data into a pre-existing supervised domain.

In summary, BluRef represents a significant advancement in image restoration, proving that unsupervised learning with dense matching references can rival supervised methods in both performance and practical applicability.