Abstract
Motion blur restoration remains a challenging task in computer vision, often requiring a balance between local texture recovery and global structural consistency. This paper proposes a novel physics-informed "Global Feature Layer" that integrates advection-diffusion equations into deep learning architectures. By modeling feature evolution through a differentiable PDE solver and employing a progressive training strategy to ensure numerical stability, the authors achieve significant performance gains (up to +1.05 dB PSNR) across state-of-the-art models with a minimal computational overhead of approximately 1%.
Introduction
Motion blur, resulting from relative motion between the camera and the scene, significantly degrades image quality and hinders downstream tasks like object detection and tracking. Traditional deep learning approaches typically rely on either CNNs, which have limited receptive fields, or Transformers, which offer global context at high computational costs. Most existing models are purely data-driven, neglecting the physical principles of blur formation. This work bridges the gap by injecting physical priors—specifically how information flows and diffuses—into neural networks to enhance restoration efficiency and effectiveness.
Methodology
PDE-based Global Feature Layer
The core contribution is a Global Feature Layer inserted at the bottleneck of a U-Net style architecture. This layer models feature evolution using an advection-diffusion equation:
$$\frac{\partial H}{\partial t} = \nabla \cdot (D\nabla H) - \nabla \cdot (\mathbf{v}H) + f(I)$$
In this formulation:
- Advection ($\mathbf{v}$): Directs information flow along image structures, facilitating the recovery of sharp edges and textures.
- Diffusion ($D$): Provides adaptive smoothing to mitigate noise and artifacts.
- Source Term ($f(I)$): Anchors the evolution process to the original degraded input features.
Differentiable Solver and Progressive Strategy
The PDE is solved using a finite-difference scheme, transformed into a series of differentiable operations. Unlike the self-attention mechanisms in Transformers, which often exhibit quadratic complexity relative to image size, the PDE solver maintains linear complexity $O(K \times H \times W)$. This allows for efficient global spatial modeling.
However, training deep networks with iterative solvers often leads to unstable gradients. To address this, the authors introduce a Progressive Integration Strategy. As shown in the figure above, the strategy involves a multi-phase training schedule where the number of iterations ($K$) increases while the time step ($\Delta t$) decreases, keeping the total integration time ($T = K \cdot \Delta t$) constant. This allows the network to stabilize with coarse temporal resolutions before refining them in later stages.
Experimental Evaluation
The proposed layer was integrated into four state-of-the-art (SOTA) models: FFTformer, NAFNet, Restormer, and Stripformer. Evaluations were conducted on the GoPro dataset (synthetic) and the RealBlur-R/J datasets (real-world RAW and JPEG).
Key Results
The PDE-enhanced models consistently outperformed their baselines across all benchmarks, with PSNR improvements ranging from +0.24 dB to +1.05 dB. Notably, the gains were most pronounced on the RealBlur-J dataset, suggesting that the physical constraints of the PDE help disentangle motion blur from JPEG compression artifacts more effectively than standard convolutions.
Crucially, this performance boost comes with negligible computational cost. As detailed in the table above, the addition of the PDE layer increases the total GMACs by only 0.6 to 2.4, representing a mere ~1% overhead across various architectures.
Conclusion
By integrating physical advection-diffusion priors into deep restoration pipelines, this work demonstrates that global spatial modeling does not require the quadratic complexity of standard self-attention. The proposed progressive training strategy is essential for the stability of such physics-informed layers. Future research may explore adaptive integration times and the application of different PDEs to tasks like dehazing or deraining.
Summary updated