Summary
Deep Unfolded BM3D turns the patch matching, collaborative filtering, and aggregation stages of BM3D into a differentiable architecture. The goal is to retain the transparency of a classical non-local prior while allowing its parameters to be learned from data.
Rather than replacing BM3D with a generic deep network, the paper keeps the algorithmic skeleton visible and trainable. That makes the model easier to reason about than a standard end-to-end denoiser while still allowing data-driven tuning of its internal operations.
Method
The approach unrolls the main BM3D stages into a sequence of differentiable modules. Patch grouping, transform-domain filtering, and aggregation are preserved conceptually, but thresholds and transforms become parameters that can be optimized from data.
Why it matters
This work extends the lab’s interest in interpretable and hybrid imaging methods. It shows how classical non-local denoising logic can survive inside a trainable model instead of being discarded when deep learning is introduced.
Highlights
- Uses algorithm unrolling rather than discarding the original BM3D structure.
- Preserves non-local self-similarity as a first-class prior.
- Aims for interpretable denoising beyond standard encoder-decoder baselines.