Improving Neural Blind Deconvolution

Abstract

The field of blind image deblurring was for a long time dominated by Maximum-A-Posteriori methods seeking the optimal pair of sharp image-blur of a suitable functional. Recently, learning-based methods, especially those based on deep convolutional neural networks, are proving effective and are receiving increasing attention by the research community. In 2020, Ren et al. proposed a deblurring method called SelfDeblur which combines the model-driven approach of traditional MAP methods and the generative power of neural nets. The method is capable of producing very high-quality results, yet it inherits some problems of MAP methods, especially possible convergence to a wrong local optimum. In this paper we propose several easy-to-implement modifications of SelfDeblur, namely suitable initialization, multiscale processing, and regularization, that improve the average performance of the original method and decrease the probability of failure.

Wiener Guided DIP for Unsupervised Blind Image Deconvolution

“…in [27] where z I and z k are random vectors sampled from U(0, 1). However, in the published code it has been suggested to remove the TV term and switch from the MSE loss to the structural similarity index measure (SSIM) after 1k iterations which is also confirmed by Kotera et al [9].…”

“…We do not perform non-blind deconvolution for Ren et al [27] and W-DIP as it has been shown that the generated image by the convolutional neural network is comparable to that of the non-blind deconvolved version [27]. We do not present any comparison to Kotera et al [9] because the code is not publicly available and the evaluation metrics used in the paper are different than the ones that are commonly used in the literature.…”

“…If no prior information is available on the shape of the k, we initialize it with a Gaussian and determine the variance of the Gaussian by looking at the PSD of an image as shown in Eqs. (9).…”

“…2. On the contrary, DIP fits faster to blurred images compared to sharp images as shown in [9]. Therefore, DIPbased blind deconvolution without a suitable prior suffers from two major issues: (i) performance fluctuations across multiple runs and initialization, (ii) sensitivity to critical hyper-parameters that are difficult to predetermine, in particular the size of the kernel to be estimated as we further elaborated in Tab.…”

“…2. Kotera et al [9] proposed to improve upon [27] by initializing the kernel generator with a Gaussian and using multi-scale processing, whereas Li et al [16] proposed an architecture change as well as sampling approach. However, both methods do not explicitly leverage the strength of DIP in removing high frequency artifacts for blind image deconvolution.…”