Multi-Observation Blind Deconvolution with an Adaptive Sparse Prior.

Area of Science:

• Computer Vision
• Image Processing
• Computational Imaging

Background:

• Multi-image blind deconvolution is challenging due to unknown blur kernels and noise.
• Existing methods often struggle with varying levels of image degradation.

Purpose of the Study:

• To develop a robust algorithm for estimating a single sharp latent image from multiple blurry and/or noisy observations.
• To jointly estimate latent images, blur kernels, and noise variances without prior degradation type knowledge.

Main Methods:

• A Bayesian-inspired penalty function links multiple observations.
• The algorithm adaptively weights observations based on their quality (concavity/sparsity).
• Joint estimation of latent image, blur kernels, and noise variances.

Main Results:

• The algorithm recovers sharp images from mixed blurry and noisy inputs.
• It automatically prioritizes higher-quality observations.
• Avoids local minima and requires no essential tuning parameters.

Conclusions:

• The proposed method effectively addresses multi-image blind deconvolution with unknown degradation types.
• Demonstrates robust performance on synthetic and real-world image data.
• Offers a significant advancement in image restoration techniques.