A Study of the Limits of Collaborative DCT-Based Image Denoising via Interpretable Neural Networks

Modern image denoising often relies on large, black-box deep neural networks that lack interpretability. Conversely, classical methods like BM3D offer clear, structured pipelines—grouping, filtering, and aggregation—but rely on handcrafted, non-differentiable operations. This paper asks: can we reformulate these structured collaborative filtering principles into a fully differentiable, trainable architecture that bridges the gap between classical interpretability and modern deep learning performance?

The authors introduce DeepBM3D, a two-stage, end-to-end trainable architecture. It replaces the fixed components of classical BM3D with learnable modules:

• Patch Grouping: Uses a differentiable attention-based mechanism (PSAL) to group similar patches, guided by lightweight convolutional feature extractors.
• Collaborative Filtering: Performs filtering in the DCT domain using learned Wiener weights, which are predicted by a small CNN.
• Refinement: Employs a two-stage pipeline where the first stage provides a basic estimate that guides the second stage, mirroring the iterative refinement strategy of classical BM3D.

The model is trained using an L1 loss across multiple noise levels, with both flexible (multi-noise) and fixed (noise-specific) configurations.

DeepBM3D consistently outperforms classical and hybrid baselines across standard benchmarks. It remains highly competitive with FFDNet, a representative deep learning denoiser, particularly at low-to-moderate noise levels. The model excels at reconstructing repetitive textures, where its non-local grouping and transform-domain filtering priors are most effective. Ablation studies confirm that the two-stage refinement and larger patch groups significantly boost performance, though the model's structural constraints make it slightly less effective than unconstrained CNNs in high-noise scenarios.