Abstract The goal of blind image deblurring is to recover a sharp image from a motion blurred one without knowing the camera motion. Current state-of-the-art methods have a remarkably good performance on images with no noise or very low noise levels. However, the noiseless assumption is not realistic considering that low light conditions are the main reason for the presence of motion blur due to requiring longer exposure times. In fact, motion blur and high to moderate noise often appear together. Most works approach this problem by first estimating the blur kernel $k$ and then deconvolving the noisy blurred image. In this work, we first show that current state-of-the-art kernel estimation methods based on the $\ell_0$ gradient prior can be adapted to handle high noise levels while keeping their efficiency. Then, we show that a fast non-blind deconvolution method can be significantly improved by first denoising the blurry image. The proposed approach yields results that are equivalent to those obtained with much more computationally demanding methods.
Introduction
Blind deblurring is a classic "chicken and egg" problem in computer vision: to deblur an image, one needs to know the blur kernel (the path the camera took while the shutter was open), but to estimate that kernel accurately, a sharp version of the image is usually required. While many algorithms work well on clean images, they often fail in real-world scenarios. Motion blur typically occurs in low-light conditions where exposure times are long, which also results in high sensor noise. Most existing algorithms mistake noise for image edges or kernel features, leading to "ringing" artifacts or completely incorrect kernel estimates. This paper addresses the challenge of separating structural information needed for deblurring from the random noise that obscures it.
Methodology
The proposed framework utilizes a two-stage approach: Kernel Estimation followed by Non-blind Deconvolution.
1. Robust Kernel Estimation
The method iterates between predicting a sharp image and estimating the blur kernel. By using an $L_0$ gradient prior, which encourages flat regions and sparse sharp edges, and increasing the regularization threshold ($\lambda_{min}$), the prior acts as a hard threshold. This allows the algorithm to ignore noise spikes while preserving the main structures necessary for kernel estimation. Furthermore, $L_1$ (sparsity) and $L_2$ (smoothness) priors are incorporated directly into the kernel optimization to ensure physical plausibility. Efficiency is maintained through a coarse-to-fine multiscale scheme with optimized "warm-starting" iterations.
2. Denoise-then-Deconvolve
Once the kernel is estimated, the final image is reconstructed. The authors demonstrate that a sequential approach is highly effective: the blurry image is first denoised using a modern CNN-based denoiser (FFDNet) and then deconvolved using a standard Total Variation (TV) regularized algorithm. This prevents the deconvolution process from amplifying residual noise into artifacts.
Results and Discussion
The proposed method was evaluated against standard benchmarks like the Levin et al. and Sun et al. datasets, as well as real-world images.
On the Levin dataset with 10% noise, the proposed method achieved a lower Root Mean Square Error (RMSE) for kernel estimation (0.151) compared to competitors like Pan et al. (0.171). In terms of image quality, the method reached a PSNR of 25.10 dB, significantly outperforming Tao et al. (20.28 dB) and Zhong et al. (23.38 dB).
Beyond accuracy, the algorithm is remarkably efficient. For a $512 \times 512$ image, it runs in 17 seconds, which is up to 10x faster than some competing methods (e.g., Pan et al. at 165s). This speed and robustness stem from the insight that the $L_0$ gradient prior effectively "ignores" noise below a certain intensity threshold when tuned correctly.
Conclusion
This work demonstrates that robust and efficient blind deblurring is possible even under high noise levels by adapting existing priors and adopting a simple "denoise-then-deconvolve" strategy. While the current implementation assumes stationary blur and does not specifically address JPEG artifacts, it provides a strong foundation for real-world applications. Future work may involve integrating neural networks trained specifically on combined blur and noise to further refine the deconvolution stage.
Summary updated