Heavy-Tailed Self-Similarity Modeling for Single Image Super Resolution.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Self-similarity in natural images is crucial for image restoration tasks.
• Existing super-resolution methods often require extensive training data.

Purpose of the Study:

• To propose a novel probabilistic model for Single Image Super Resolution (SISR) based on image patch similarity.
• To develop a Variational Bayes algorithm for super-resolution that does not require training.

Main Methods:

• A probabilistic model leveraging heavy-tailed distributions for quantifying image patch similarity.
• Derivation of a Variational Bayes algorithm for SISR.
• Mathematical proof demonstrating the algorithm as an extension and improvement over Non-Local Means (NLM).

Main Results:

• The proposed algorithm achieves performance comparable to trained deep learning models for rural images (SSIM).
• It outperforms Non-Local Means (NLM) and is favored in perceptual quality evaluations.
• Performance is inferior to deep learning methods for urban images (MSE).

Conclusions:

• The novel approach offers a competitive, training-free alternative for SISR, especially for specific image types.
• Highlights the limitations of Mean-Squared-Error (MSE) as a sole metric for image quality assessment.
• Suggests the potential of patch similarity models in image restoration.