Adversarial Training for Solving Inverse Problems in Image Processing.

Area of Science:

• Mathematics
• Computer Science
• Image Processing

Background:

• Inverse problems estimate source data from limited observations.
• Current deep learning methods often ignore underlying physics in image processing inverse problems.
• Pixel-wise regression is a common but limited approach.

Purpose of the Study:

• To propose a novel framework for solving inverse problems in image processing.
• To re-examine inverse problems from a different viewpoint than direct pixel-wise regression.
• To improve the estimation of source data and operation parameters.

Main Methods:

• A deep neural network is trained to estimate degradation parameters using an adversarial training paradigm.
• The framework introduces adversarial constraints to the parameter space.
• The method explores scenarios where pair-wise supervision may not be required.

Main Results:

• The proposed method demonstrates effectiveness across various real-world image processing tasks.
• Applications include image denoising, deraining, shadow removal, and illumination correction.
• The approach also shows promise in underdetermined blind source separation for images and speech.

Conclusions:

• The novel adversarial framework offers an improved approach to inverse problems in image processing.
• Estimating degradation parameters adversarially enhances solution quality.
• The method's flexibility in supervision requirements broadens its applicability.