Summary
The paper revisits total-variation denoising and adds a compact CNN that predicts a spatial lambda map. Instead of one global regularization strength, the method adapts smoothing pressure to local anatomy while keeping the method interpretable.
This gives the model a middle ground between purely handcrafted reconstruction priors and unconstrained black-box learning. The classical total-variation structure remains visible, but its strength becomes data-adaptive across the image.
Method
A lightweight CNN predicts a pixel-wise regularization map that modulates how strongly TV smoothing is applied in different regions. Structured areas can be preserved more carefully while flatter regions can be denoised more aggressively, all within a formulation that still looks like a regularized reconstruction method rather than a generic image translator.
Why it matters
This paper is useful because it treats interpretability as a design goal rather than a side effect. It shows how a familiar imaging prior can be retained while still benefiting from learned spatial adaptivity.
Highlights
- Blends classical model-based denoising with learned spatial adaptation.
- Preserves the convex structure and intuition of total variation.
- Targets compact, physically interpretable priors instead of purely black-box behavior.