Architecture-Agnostic Untrained Network Priors for Image Reconstruction with Frequency Regularization.

Area of Science:

• Computer Vision
• Machine Learning
• Medical Imaging

Background:

• Untrained networks leverage deep image priors for image recovery without training sets.
• Success is linked to spectral bias in network architectures, acting as implicit regularization.
• Current methods face challenges like architectural over-specification, overfitting, and long optimization times.

Purpose of the Study:

• To develop efficient, architecture-agnostic techniques to directly modulate the spectral bias of network priors.
• To address challenges of overfitting and lengthy optimization in untrained network applications.
• To enable the use of compact models with reduced runtime for superior or comparable performance.

Main Methods:

• Implementation of bandwidth-constrained input.
• Integration of bandwidth-controllable upsamplers.
• Application of Lipschitz-regularized convolutional layers.

Main Results:

• Demonstrated reduction in overfitting for underperforming architectures.
• Closed performance gaps between different network architectures with minimal code changes.
• Achieved comparable or superior performance using more compact models with reduced runtime.
• Successfully applied to inpainting-like MRI reconstruction tasks.

Conclusions:

• Architectural biases, overfitting, and runtime issues in untrained network priors can be addressed simultaneously without architectural modifications.
• Proposed techniques offer efficient and practical solutions for enhancing untrained network performance.
• The findings pave the way for more accessible and efficient deep learning-based image reconstruction.