Abstract:Conventional deconvolution methods utilize hand-crafted image priors to constrain the optimization. While deep-learning-based methods have simplified the optimization by end-to-end training, they fail to generalize well to blurs unseen in the training dataset. Thus, training image-specific models is important for higher generalization. Deep image prior (DIP) provides an approach to optimize the weights of a randomly initialized network with a single degraded image by maximum a posteriori (MAP), which shows that the architecture of a network can serve as the hand-crafted image prior. Different from the conventional hand-crafted image priors that are statistically obtained, it is hard to find a proper network architecture because the relationship between images and their corresponding network architectures is unclear. As a result, the network architecture cannot provide enough constraint for the latent sharp image. This paper proposes a new variational deep image prior (VDIP) for blind image deconvolution, which exploits additive hand-crafted image priors on latent sharp images and approximates a distribution for each pixel to avoid suboptimal solutions. Our mathematical analysis shows that the proposed method can better constrain the optimization. The experimental results further demonstrate that the generated images have better quality than that of the original DIP on benchmark datasets. The source code of our VDIP is available at <a class="link-external link-https" href="https://github.com/Dong-Huo/VDIP-Deconvolution" rel="external noopener nofollow">this https URL</a>.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is to perform blind image deconvolution on a single blurry image, that is, to restore the underlying clear image without knowing the blur kernel. Traditional methods usually use hand - designed image priors to constrain the optimization process, while deep - learning - based methods, although simplifying the optimization process, have poor generalization ability for blur types not present in the training dataset. Therefore, training image - specific models is crucial for improving generalization performance. ### Main contributions of the paper 1. **Proposed a new variational deep image prior (VDIP)**: By combining the deep image prior (DIP) and the variational Bayes method, VDIP can better constrain the optimization process and avoid the local optimum problem in the sparse maximum a posteriori probability (Sparse MAP) method. 2. **Provided a complete derivation and mathematical analysis of the final loss function**: It is proved that the proposed method has a better constraining effect in the optimization process than the traditional DIP method. 3. **Experimental results show**: The quantitative results of VDIP on the benchmark dataset and the quality of the generated clear images are significantly better than those of the traditional DIP method. ### Main technical details - **Super - Gaussian distribution**: Traditional image priors can be represented as super - Gaussian distributions. By introducing variational parameters, non - convex optimization problems are transformed into convex optimization problems. - **Variational inference**: Using the variational Bayes method, the posterior distribution is approximated by minimizing the KL divergence, thus avoiding the local optimum problem in the sparse MAP method. - **Variational deep image prior**: Two networks are used to learn the distributions of the underlying clear image and the blur kernel respectively. The expectation term is approximated by Monte Carlo estimation, making the optimization objective differentiable. ### Experimental results - **Quantitative comparison**: On the synthetic dataset, VDIP is significantly better than the traditional DIP and other methods in terms of PSNR and SSIM metrics. - **Blur kernel recovery error**: VDIP also performs well in the recovery accuracy of the blur kernel, especially when dealing with saturated images. ### Summary This paper improves the traditional deep image prior method by introducing the variational Bayes method, solves the local optimum problem in blind image deconvolution, and improves the quality of the restored clear image and the recovery accuracy of the blur kernel.

Blind Image Deconvolution Using Variational Deep Image Prior

Multi-frame Blind Deconvolution Using Sparse Priors

Non-blind image deconvolution using natural image gradient prior

VDIP-TGV: Blind Image Deconvolution via Variational Deep Image Prior Empowered by Total Generalized Variation

Blind Deconvolution Using an Improved L0 Sparse Representation

High quality non-blind image deconvolution using the Fields of Experts prior

Blind Image Deblurring Via Deep Discriminative Priors

Learning Deep Optimizer For Blind Image Deconvolution

A Deep Optimization Approach for Image Deconvolution.

Blind Image Deconvolution using Deep Generative Priors

Blind image deblurring via hybrid deep priors modeling

A Deep Variational Bayesian Framework for Blind Image Deblurring

Deep Prior-Based Blind Image Deblurring

Blind Image Deblurring Using Adaptive Priors.

Neural blind deconvolution using deep priors

Blind Image Deconvolution by Generative-based Kernel Prior and Initializer via Latent Encoding

Image Deconvolution with Deep Image and Kernel Priors

Blind Image Deconvolution by Automatic Gradient Activation.

Blind Image Deblurring Using Elastic-Net Based Rank Prior

Discriminative learning of iteration-wise priors for blind deconvolution