Real-World Efficient Blind Motion Deblurring via Blur Pixel Discretization

TL;DR

The paper tackles the challenge of real-world blind motion deblurring on high-resolution imagery by reframing the regression problem as a two-stage process: first discretize per-pixel blur into a small set of classes via a blur segmentation map, then perform a discrete-to-continuous conversion to recover the sharp image. Central to the approach is leveraging a logarithmic Fourier cepstrum to relate blur and sharp images and incorporating a latent sharp image to constrain the solution space, enabling a light-weight model to achieve competitive results. The proposed SegDeblur variants demonstrate strong PSNR/SSIM on RealBlur, RSBlur, and ReLoBlur datasets at significantly lower MACs than kernel-free baselines and outperform kernel-based priors with far less complexity, achieving practical on-device deployment potential. The work suggests that this discretization-prior framework can extend to video deblurring and related blur types, offering a pathway to efficient, real-time deblurring in mobile and embedded settings.

Abstract

As recent advances in mobile camera technology have enabled the capability to capture high-resolution images, such as 4K images, the demand for an efficient deblurring model handling large motion has increased. In this paper, we discover that the image residual errors, i.e., blur-sharp pixel differences, can be grouped into some categories according to their motion blur type and how complex their neighboring pixels are. Inspired by this, we decompose the deblurring (regression) task into blur pixel discretization (pixel-level blur classification) and discrete-to-continuous conversion (regression with blur class map) tasks. Specifically, we generate the discretized image residual errors by identifying the blur pixels and then transform them to a continuous form, which is computationally more efficient than naively solving the original regression problem with continuous values. Here, we found that the discretization result, i.e., blur segmentation map, remarkably exhibits visual similarity with the image residual errors. As a result, our efficient model shows comparable performance to state-of-the-art methods in realistic benchmarks, while our method is up to 10 times computationally more efficient.

Figures (28)

• Figure 1: Blind motion deblurring via blur pixel discretization. The image residual error, i.e., blur-sharp pixel differences is estimated by our blur pixel discretization and discrete-to-continuous conversion, whose result is used to produce final deblurred images.
• Figure 2: Visual comparison in the object motion blur (above) and uniform blur (below): (a) Sharp image, (b) Blur image, (c) Image residual error, and (d) Blur segmentation map. We utilize three classes and the colors in blur segmentation map indicate classes.
• Figure 3: Visual comparison results on the efficient models: (a) Blur image, (b) NAFNet nafnet, and (c) Proposed method. The computational cost of both models is around $16$ GMACs. NAFNet is vulnerable to distortions since it has no blur class information, whereas our method produces more natural deblurring results.
• Figure 4: Our network architectures. The proposed method consists of two stages and each stage is shaded in different colors. In the first stage, we train our blur pixel discretizer and kernel estimator. In the second stage, we proceed to train our D2C converter with the frozen blur pixel discretizer. LogFT indicates the logarithmic fourier transform and IFTExp means the inverse logarithmic fourier transform.
• Figure 5: Visual comparison results on RSBlur rsblur. We compare our SegDeblur-S ($14.44$ GMACs) with FFTFormer-16 fftformer ($16.41$ GMACs) and NAFNet-32 nafnet ($16.25$ GMACs). Note that all methods are trained with RSBlur.