Multi-scale plug-and-play energy framework for inverse problems

Motivation: Unrolled algorithms provide high quality image reconstruction. However, their training is memory-intensive and is sensitive to forward model mismatches. Goal(s): To develop a memory-efficient plug-and-play algorithm, whose performance is comparable to unrolled algorithms and can be used with arbitrary forward models. Approach: We propose a memory-efficient energy-based multi-scale framework. We model the negative log prior with different smoothnesses using Convolutional Neural Networks (CNN). This approach enables us to relax the constraints on the CNN, while the multi-scale strategy improves the convergence to the global minimum. Results: The enhancements improves performance, making it comparable to end-to-end methods, while being robust to model mismatch. Impact: The proposed framework is memory-efficient compared to unrolled algorithms, paving the way for its usage in large-dimensional inverse problems. Its flexibility enables recovery of images with arbitrary forward operators.