Abstract
The goal of dynamic scene deblurring is to remove the motion blur in a given image. Typical learning-based approaches implement their solutions by minimizing the L1 or L2 distance between the output and the reference sharp image. Recent attempts adopt visual recognition features in training to improve the perceptual quality. However, those features are primarily designed to capture high-level contexts rather than low-level structures such as blurriness. Instead, we propose a more direct way to make images sharper by exploiting the inverse task of deblurring, namely, reblurring. Reblurring amplifies the remaining blur to rebuild the original blur, however, a well-deblurred clean image with zero-magnitude blur is hard to reblur. Thus, we design two types of reblurring loss functions for better deblurring. The supervised reblurring loss at training stage compares the amplified blur between the deblurred and the sharp images. The self-supervised reblurring loss at inference stage inspects if there noticeable blur remains in the deblurred. Our experimental results on large-scale benchmarks and real images demonstrate the effectiveness of the reblurring losses in improving the perceptual quality of the deblurred images in terms of NIQE and LPIPS scores as well as visual sharpness.
Introduction
Dynamic scene deblurring—removing motion blur from images where both the camera and objects might be moving—is a classic ill-posed problem. Standard deep learning approaches usually minimize pixel-wise distance (L1 or L2 loss) between the model's output and a sharp ground-truth image. However, this often leads to "regression-to-the-mean," resulting in outputs that look safe but slightly blurry. While perceptual losses and GANs help improve realism, they often focus on high-level semantic features rather than low-level structural properties like edge sharpness. Consequently, even state-of-the-art models often leave behind "blur footprints."
Methodology: The Reblurring Principle
The core insight of this work is that clean images are hard to reblur. If an image is perfectly sharp, a reblurring module has no motion information to latch onto and amplify. Conversely, if the image still contains residual blur, that blur can be easily "stretched" back into the original blurry state.
Reblurring Loss Functions
The researchers implemented this via two main components:
- Supervised Reblurring Loss (Training): A deblurring module ($\mathcal{M}_D$) and a reblurring module ($\mathcal{M}_R$) are trained together. The reblurring module attempts to reconstruct the original blurry image from the deblurred output. The loss function encourages the deblurrer to produce an image that makes the reblurrer's job as difficult as possible compared to the ground-truth sharp image.
- Self-Supervised Test-Time Adaptation (TTA): At inference time, the model's weights can be iteratively updated for a specific input image. By minimizing a self-supervised reblurring loss—essentially making the image "harder to reblur"—the model sharpens the specific textures and edges of that unique scene.
Experimental Results
The models were evaluated on standard large-scale benchmarks including GOPRO and REDS, as well as real-world blurry photos.
Qualitative Comparison
Visually, the results show much crisper edges and finer details compared to previous state-of-the-art methods. The reblurring loss acts as a task-specific perceptual guide that preserves structural integrity better than generic semantic features.
As illustrated in the figure above, the proposed method (e) achieves superior sharpness compared to SRN (b), SE-Sharing (c), and DeblurGANv2 (d). The bottom row demonstrates the "reblurring" concept: when the deblurred images are reblurred (g-j), the proposed method (j) shows the least amount of residual blur amplification, indicating a cleaner initial deblurring result.
Quantitative Gains and Trade-offs
The method consistently improved perceptual metrics across different architectures (U-Net, SRN, DHN), achieving better LPIPS (Learned Perceptual Image Patch Similarity) and NIQE (Natural Image Quality Evaluator) scores. Interestingly, while perceptual sharpness increased, the PSNR (Peak Signal-to-Noise Ratio) sometimes decreased slightly. This is a known phenomenon in image restoration where pixel-perfect alignment often looks blurrier to the human eye than a slightly shifted but sharp edge.
Conclusion and Limitations
This paper provides a powerful new perspective: the best way to solve a problem is often to look at how hard it is to undo the solution. While the Test-Time Adaptation (TTA) offers significant quality boosts, it comes with a computational cost that may limit real-time applications. Future work may focus on reducing this overhead and improving color stability during the adaptation process.
Summary updated