Graph-Based Blind Image Deblurring From a Single Photograph.

Area of Science:

• Computer Vision
• Image Processing
• Graph Signal Processing

Background:

• Blind image deblurring is an ill-posed problem requiring blur kernel estimation and deconvolution.
• Existing methods often struggle with accuracy and efficiency in kernel estimation.

Purpose of the Study:

• To develop a novel graph-based algorithm for blind image deblurring.
• To accurately estimate blur kernels using skeleton images and a new prior.
• To improve image restoration quality and robustness.

Main Methods:

• Interpreting image patches as signals on weighted graphs.
• Utilizing skeleton images with bi-modal edge weight distribution for kernel estimation.
• Designing a reweighted graph total variation (RGTV) prior for promoting bi-modal distributions.
• Developing an efficient alternating optimization algorithm leveraging graph spectral filtering.
• Proposing an accelerated strategy for blind Gaussian deblurring.

Main Results:

• The proposed RGTV prior enables a graph spectral filtering interpretation, offering robustness to noise and blur.
• The algorithm efficiently estimates blur kernels and restores latent sharp images.
• Experimental results show superior quantitative and qualitative performance compared to state-of-the-art methods.

Conclusions:

• The graph-based approach with RGTV provides an effective solution for blind image deblurring.
• The method demonstrates significant improvements in image restoration accuracy and efficiency.
• This work offers a new perspective on blind deblurring through graph spectral analysis.