Chapter 20: Post-Processing and Model-Based Networks

Learning Objectives

• Implement the matched-filter-to-U-Net pipeline and analyze why it fails for structured sensing matrices
• Explain the sidelobe corruption problem when the Gram matrix $\mathbf{G} = \mathbf{A}^{H}\mathbf{A}$ has strong off-diagonal structure
• Design data-consistency layers and MoDL-style alternating architectures for inverse problems
• Build physics-informed post-processing networks that condition on the sensing geometry
• Select training strategies: loss functions (MSE, perceptual, adversarial), data augmentation, transfer learning