Deep Gaussian Scale Mixture Prior for Image Reconstruction.

Area of Science:

• Computer Vision
• Machine Learning
• Signal Processing

Background:

• Image reconstruction from incomplete data is challenging.
• Traditional methods struggle with fine details due to limited prior representations.
• Deep learning offers improved performance but often lacks transparency.

Purpose of the Study:

• To develop a novel, transparent image reconstruction method.
• To improve the recovery of fine image details from partial observations.
• To leverage learned priors for enhanced reconstruction accuracy.

Main Methods:

• Proposed a Maximum a Posteriori (MAP) estimation framework incorporating learned Gaussian Scale Mixture (GSM) priors.
• Developed a deep network to learn both means and variances of GSM models, unlike prior methods focusing only on means.
• Utilized a Swin Transformer variant for learning long-range image dependencies and jointly optimized all parameters end-to-end.

Main Results:

• The proposed method successfully reconstructs images with finer details compared to conventional approaches.
• Learned GSM priors with both means and variances significantly improve representation capabilities.
• Demonstrated superior performance on spectral compressive imaging and image super-resolution tasks using simulations and real data.

Conclusions:

• The novel MAP-based method with learned GSM priors offers a more effective approach to image reconstruction.
• Incorporating learned variances and long-range dependencies via Swin Transformers enhances reconstruction quality.
• The end-to-end trained deep network provides a transparent and powerful solution for complex imaging problems.