Learning Deep Gradient Descent Optimization for Image Deconvolution.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Non-blind deconvolution is crucial for image deblurring but challenging due to the ill-posed nature of inverse problems.
• Existing learning-based methods offer high restoration quality but lack practicality due to static model designs and reliance on known noise levels.

Purpose of the Study:

• To bridge the gap between optimization-based and learning-based deblurring methods.
• To develop a universal gradient descent optimizer for practical blind image deblurring.

Main Methods:

• Proposed a recurrent gradient descent network (RGDN) integrating deep neural networks into a gradient descent scheme.
• Employed a hyperparameter-free update unit based on a convolutional neural network for generating updates.
• Trained the RGDN on diverse examples for recursive supervision, learning an implicit image prior and a universal update rule.

Main Results:

• The RGDN demonstrated strong interpretability and high generalization capabilities.
• Extensive experiments on synthetic and real-world images confirmed the method's effectiveness and robustness.
• The learned optimizer improved the quality of diverse degenerated observations through repeated application.

Conclusions:

• The proposed deep optimization method offers a practical and effective solution for real-world image deblurring.
• The RGDN learns a universal update rule, enhancing its applicability across various deblurring scenarios.
• This approach provides a more flexible and powerful alternative to existing deblurring techniques.