CAR-Net: A Cascade Refinement Network for Rotational Motion Deblurring under Angle Information Uncertainty

TL;DR

Rotational motion deblurring under uncertain angle information is addressed with CAR-Net, a cascade refinement network that first performs a PCS-based frequency inversion and then iteratively refines the result while optionally correcting the blur angle via an Angle Detection Module. The method combines physics-based modeling with data-driven refinements, yielding a robust, efficient solution for semi-blind RMB. Through extensive ablations and real-world testing, CAR-Net-AD demonstrates superior PSNR/SSIM and practical speed compared to the state-of-the-art PCS-RMD, while revealing insights on loss weighting and training data requirements. The work highlights the value of decoupling parameter correction from image refinement in a modular, cascaded architecture with potential applicability to other parameter-uncertainty restoration problems.

Abstract

We propose a new neural network architecture called CAR-net (CAscade Refinement Network) to deblur images that are subject to rotational motion blur. Our architecture is specifically designed for the semi-blind scenarios where only noisy information of the rotational motion blur angle is available. The core of our approach is progressive refinement process that starts with an initial deblurred estimate obtained from frequency-domain inversion; A series of refinement stages take the current deblurred image to predict and apply residual correction to the current estimate, progressively suppressing artifacts and restoring fine details. To handle parameter uncertainty, our architecture accommodates an optional angle detection module which can be trained end-to-end with refinement modules. We provide a detailed description of our architecture and illustrate its efficiency through experiments using both synthetic and real-life images. Our code and model as well as the links to the datasets are available at https://github.com/tony123105/CAR-Net

Figures (4)

• Figure 1: The architecture of our Baseline Model. The input polar image $g_\mathrm{p}$ is first processed by the Frequency Inversion Module (INV) to produce an initial estimate $f_0$. This estimate is then passed to the Progressive Spatial Refinement Module (RS), which iteratively remove artifacts and restores details. Green and gray blocks denote trainable and non-trainable modules, respectively.
• Figure 2: The architecture of our Full Model which extends the baseline by inserting the Angle Detection Module (AD) into a feedback loop. The AD accept the initial deblurred image ($f^{-1}$) and the initial angle ($\theta_\mathrm{initial}$) to predict a corrected angle ($\theta_\mathrm{corrected}$). The new angle is used to perform a more accurate inversion before the final refinement stages. Green and gray blocks denote trainable and non-trainable modules, respectively.
• Figure 3: The architectural components of our proposed model. (a) The module for correcting the blur angle. (b) The residual block. (c) The iterative refinement stage. (d) The final refinement block. Module shown with $\Delta$ represents where the feature map is downsampled
• Figure 4: A visual comparison of deblurring results on an image from the Real-World test set. (Noise uncertainty level $\sigma=5$, ground truth angle $\theta_{\mathrm{GT}}=5.41\degree$, noisy initial angle $\theta_{\mathrm{initial}}=8.98\degree$).