Blind Deconvolution With Model Discrepancies.

Area of Science:

• Computer Vision
• Image Processing
• Signal Processing

Background:

• Blind deconvolution is an ill-posed problem requiring simultaneous blur and image estimation.
• Existing methods often fail when the convolution model is violated in parts of an image.
• Variational Bayesian inference methods are prominent but have limitations.

Purpose of the Study:

• To develop a robust blind deconvolution method that addresses limitations of current approaches.
• To improve image and blur estimation in the presence of real-world degradations.
• To extend the applicability of blind deconvolution to challenging scenarios.

Main Methods:

• Utilizing variational Bayesian inference with an automatic relevance determination (ARD) prior.
• Applying the ARD prior to noise, image, and blur components.
• Developing a method that handles discrepancies in the convolution model.

Main Results:

• The ARD prior enforces non-negativity of the blur kernel and favors sharper images.
• The method effectively handles non-Gaussian noise, common in real-world images.
• Demonstrated robustness to discrepancies in the convolution model.

Conclusions:

• The proposed blind deconvolution method enhances robustness in real-world applications.
• It extends the applicability of blind deconvolution to images with camera motion and focus issues.
• The ARD prior offers significant advantages for handling complex image degradation scenarios.