Abstract
Blind Image Deconvolution (BID) is a fundamental yet ill-posed problem in image processing, where both the sharp image and the blur kernel must be estimated from a single degraded observation. This paper introduces a robust, fully automated Bayesian framework for BID. The approach leverages non-convex $l_p$ sparse priors ($0 < p < 1$) to model the heavy-tailed distribution of natural image gradients and Total Variation (TV) regularization for the blur kernel. Optimization is achieved through the Majorization-Minimization (MM) principle, which facilitates the handling of non-convex priors. A significant contribution is the automatic estimation of all hyperparameters, eliminating the need for manual tuning. Experimental results on synthetic and real-world datasets demonstrate that the proposed method achieves superior performance compared to current state-of-the-art techniques, particularly in handling complex camera-shake blurs.
Introduction
Blind Image Deconvolution (BID) is a classic "one equation, two unknowns" problem. Given a blurry, noisy observation $y$, the goal is to recover both the original sharp image $x$ and the blur kernel $H$ (the point spread function). Mathematically, this is represented as $y = x * H + n$, where $n$ is noise. This problem is inherently ill-posed, as infinitely many pairs of $(x, H)$ can produce the same $y$.
Most existing solutions struggle with two primary challenges: prior selection and parameter tuning. Natural images possess specific statistical properties—gradients are usually zero but occasionally very large—which standard Gaussian models fail to capture. Furthermore, many algorithms require users to manually adjust regularization parameters, making them impractical for automated systems or general users.
Methodology
The proposed framework treats the image, blur, and noise as random variables within a Bayesian context.
Sparse Priors and Regularization
The "secret sauce" of this method is the use of a non-convex $l_p$ quasi-norm ($0 < p < 1$) for image gradients. Specifically, $p=4/5$ was found to be highly effective. This prior allows the model to be more aggressive in preserving sharp edges while suppressing noise in flat regions compared to standard $l_1$ (Laplacian) or $l_2$ (Gaussian) priors. For the blur kernel, a Total Variation (TV) prior is utilized to ensure spatial coherence.
Optimization and Automation
Because the $l_p$ norm is non-convex, the authors employ the Majorization-Minimization (MM) principle. This iterative process finds a convex "bound" for the non-convex function, solves it, and repeats, transforming a complex optimization into a series of manageable steps.
Crucially, the model automatically estimates hyperparameters—such as noise variance and prior weights—simultaneously with the image and blur. To avoid the "delta" trivial solution (where the algorithm assumes the blurry image is the sharp one and the kernel is a single dot), a multi-scale implementation is used, processing the image through a pyramid of resolutions.
Experimental Setup
The researchers evaluated the algorithm using both synthetic and real-world data. Synthetic benchmarks included standard images like the Cameraman and Satellite, degraded by five distinct non-parametric motion blurs with added Gaussian noise.
Real-world evaluation was conducted on a dataset of 32 blurred photographs representing authentic "camera shake" scenarios, where ground truth data was available for comparison.
Results and Discussion
Quantitative analysis on the 32-image real-world dataset showed that the proposed algorithm (ALG) yielded the smallest Sum of Squared Errors (SSE) in 66% of cases. Using the Error Ratio (ER) metric—where 1.0 represents the performance of non-blind deconvolution—the method achieved an ER below 1.5 in 78% of cases, significantly outperforming competitors like Cho et al. (53%).
Qualitatively, the method produces sharper restorations with fewer "ringing" artifacts (halos around edges). The automation of hyperparameter estimation makes the algorithm robust across diverse image types without requiring heuristic re-weighting.
Conclusion
While the algorithm is highly effective, it currently requires a pre-specified maximum size for the blur kernel support. Future work may focus on determining kernel size automatically, reducing the computational cost of the iterative MM approach, and extending noise modeling to handle Poisson or sensor-specific noise patterns common in low-light photography.
Summary updated