Simultaneous Fidelity and Regularization Learning for Image Restoration.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Traditional non-blind image restoration methods rely on precise degradation models, which are often unavailable or inaccurate.
• Partially known or inaccurately modeled degradations lead to suboptimal image restoration, as seen in rain streak removal and deconvolution with uncertain blur kernels.

Purpose of the Study:

• To develop a principled algorithm for image restoration that accommodates partially known or inaccurate degradation models.
• To address the limitations of existing methods in tasks like rain streak removal and blind deconvolution.

Main Methods:

• Proposes a maximum a posteriori (MAP) framework to model the complex, spatially dependent residuals from inaccurate degradations.
• Introduces a task-driven approach to parameterize and learn the fidelity term using degraded and ground-truth image pairs.
• Develops a simultaneous learning model for both fidelity and regularization terms.

Main Results:

• Demonstrates significant effectiveness in image deconvolution with inaccurate blur kernels.
• Shows successful application in deconvolution with multiple degradations.
• Achieves effective rain streak removal, addressing challenges in modeling rain streaks and their composition with scene layers.

Conclusions:

• The proposed algorithm offers a robust solution for image restoration problems with imperfect degradation models.
• Simultaneous learning of fidelity and regularization terms enhances restoration quality.
• The method proves versatile across various challenging image restoration tasks.