Deep Variational Network Toward Blind Image Restoration

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Blind image restoration (IR) is a challenging computer vision problem.
• Existing methods include classical model-based and deep learning (DL) approaches, each with limitations.

Purpose of the Study:

• To propose a novel blind image restoration method integrating advantages of both model-based and DL techniques.
• To develop a unified framework for joint degradation estimation and image restoration.

Main Methods:

• Constructed a general Bayesian generative model for blind IR, explicitly defining the degradation process.
• Employed a pixel-wise non-i.i.d. Gaussian distribution for flexible image noise modeling.
• Designed a variational inference algorithm with deep neural networks for posterior distribution parameterization.

Main Results:

• The proposed method achieves superior performance in image denoising and super-resolution compared to state-of-the-art methods.
• The unified framework effectively integrates degradation estimation and image restoration.
• The flexible noise modeling handles complex image degradation types.

Conclusions:

• The novel Bayesian generative model with variational inference offers an effective approach to blind image restoration.
• The method demonstrates significant improvements over existing techniques in key IR tasks.
• This integrated framework advances the field of computer vision for image quality enhancement.