Untrained Neural Network Priors for Inverse Imaging Problems: A Survey

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Traditional analytical methods for inverse imaging problems rely on explicit problem definitions and engineered domain knowledge.
• Machine learning (ML), particularly deep learning (DL), has emerged as a powerful alternative, often outperforming traditional approaches.
• DL models typically require large datasets to learn solutions, lacking explicit prior knowledge.

Purpose of the Study:

• To provide a comprehensive review of untrained neural network prior (UNNP) methods for inverse imaging tasks.
• To explore various applications and variants of UNNP in image restoration, inpainting, and other related problems.
• To identify and highlight open research challenges and future directions in the field of UNNP.

Main Methods:

• Reviewing existing literature on deep learning applications for inverse imaging problems.
• Analyzing the paradigm of training deep models with single images (UNNP).
• Categorizing and discussing different UNNP variants and their applications.

Main Results:

• UNNP has shown significant potential in solving various inverse imaging tasks using single-image training.
• Numerous applications and modifications of the UNNP approach have been proposed by researchers.
• The review consolidates current UNNP research and identifies areas needing further investigation.

Conclusions:

• Untrained neural network priors represent a significant advancement in solving inverse imaging problems.
• Further research is needed to fully explore the capabilities and limitations of UNNP methods.
• UNNP offers a promising avenue for future developments in image restoration, inpainting, and beyond.