Neural Blind Deconvolution Using Deep Priors (1908.02197v2)

Abstract: Blind deconvolution is a classical yet challenging low-level vision problem with many real-world applications. Traditional maximum a posterior (MAP) based methods rely heavily on fixed and handcrafted priors that certainly are insufficient in characterizing clean images and blur kernels, and usually adopt specially designed alternating minimization to avoid trivial solution. In contrast, existing deep motion deblurring networks learn from massive training images the mapping to clean image or blur kernel, but are limited in handling various complex and large size blur kernels. To connect MAP and deep models, we in this paper present two generative networks for respectively modeling the deep priors of clean image and blur kernel, and propose an unconstrained neural optimization solution to blind deconvolution. In particular, we adopt an asymmetric Autoencoder with skip connections for generating latent clean image, and a fully-connected network (FCN) for generating blur kernel. Moreover, the SoftMax nonlinearity is applied to the output layer of FCN to meet the non-negative and equality constraints. The process of neural optimization can be explained as a kind of "zero-shot" self-supervised learning of the generative networks, and thus our proposed method is dubbed SelfDeblur. Experimental results show that our SelfDeblur can achieve notable quantitative gains as well as more visually plausible deblurring results in comparison to state-of-the-art blind deconvolution methods on benchmark datasets and real-world blurry images. The source code is available at https://github.com/csdwren/SelfDeblur

An Analysis of "Neural Blind Deconvolution Using Deep Priors"

The paper, "Neural Blind Deconvolution Using Deep Priors," addresses a pivotal challenge in low-level vision problems—blind deconvolution—which entails the reconstruction of a clear image and estimation of blur kernel from a single blurred image. Traditional Maximum a Posterior (MAP) based methods, relying heavily on handcrafted priors and specialized optimization techniques, have made significant advancements. However, they face limitations in handling complex and large blur kernels. Deep learning approaches also offer promise through learned mappings from extensive training data but struggle with diverse blur scenarios.

The authors present a novel solution that integrates MAP with deep learning, proposing a neural blind deconvolution approach utilizing deep priors. Their contribution involves designing two generative networks: an asymmetric Autoencoder with skip connections to model clean image priors and a fully-connected network (FCN) with constraints for the blur kernel. This method, termed SelfDeblur, optimizes generative networks through a zero-shot self-supervised learning framework to address blind deconvolution problems.

Methodology and Contributions

SelfDeblur innovatively combines generative networks to replace handcrafted image and kernel priors in solving the deconvolution problem. The authors employ:

1. Deep Image Prior (DIP): The proposed asymmetric Autoencoder effectively models image priors, leveraging DIP's capability to capture low-level image statistics without training data.
2. Fully-Connected Network: For blur kernel estimation, a simple FCN enhanced by SoftMax nonlinearity ensures non-negativity and normalization, crucial for kernel suitability.
3. Unconstrained Neural Optimization: The joint optimization framework fosters synergy between the network models, overcoming the constraints typically managed through MAP-based alternating minimization.

Numerical Results and Performance

Their empirical evaluations on benchmark datasets demonstrate substantial improvements. Quantitatively, SelfDeblur achieves higher PSNR and SSIM metrics compared to state-of-the-art MAP-based methods, validating its efficacy. Importantly, substantial recovery of visual details is highlighted, showcasing practical applicability in real-world images plagued with motion blur. The integration of neural architectures in deconvolution paves a path to optimize both kernel estimation and image restoration, diminishing the need for additional non-blind deconvolution processes.

Implications and Future Directions

This work underscores the potential of neural networks in enhancing traditional image restoration frameworks. By effectively capturing and exploiting image and kernel priors through learned structures, the approach heralds advancements in tackling various deblurring applications and possibly extending to other inverse problems in image processing.

Future work could delve into mitigating the computational demands posed by extensive neural network training. Additionally, exploring more robust models adept at varying noise levels and real-world image conditions could extend SelfDeblur's applicability. Innovations that integrate reinforcement learning or attention mechanisms might further refine deconvolution solutions.

In conclusion, this paper makes a noteworthy contribution, bridging the gap between traditional MAP methods and deep learning frameworks, offering a compelling alternative to blind deconvolution challenges with promising implications for advancements in AI-driven image restoration technologies.