Training Very Deep CNNs for General Non-Blind Deconvolution.

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Non-blind image deconvolution is inherently ill-posed due to noise and blur.
• Existing methods require specific models for different kernels and noise levels, leading to inefficiency.
• Current techniques often produce residuals highly correlated with image content, kernel, and noise.

Purpose of the Study:

• To develop a single, generalizable model for non-blind image deconvolution.
• To address varying blur kernels and noise levels within one framework.
• To improve computational efficiency and reduce model parameter redundancy.

Main Methods:

• Proposed a very deep convolutional neural network (CNN) architecture.
• Employed a residual learning strategy, predicting the difference between a pre-deconvolved and sharp image.
• Trained a single model to handle diverse blur kernels and noise intensities.

Main Results:

• The residual learning approach facilitates training a single model for various deconvolution scenarios.
• Quantitative evaluations confirm the model's practical applicability across different blur kernels.
• Achieved state-of-the-art performance on synthesized blurry images.

Conclusions:

• The proposed deep CNN with residual learning offers an effective and efficient solution for general non-blind deconvolution.
• This approach simplifies deconvolution tasks by eliminating the need for kernel-specific models.
• The model demonstrates significant potential for real-world image restoration applications.