Normalization Equivariance Achieved via Input-Output Wrapper for Any Backbone
A new method called WNE (Wrapper Normalization Equivariance) enforces normalization equivariance—robustness to global contrast and brightness changes—on any image-to-image backbone without modifying internal layers. The key insight is that a function is normalization equivariant if and only if it can be factored as normalize-process-denormalize. This allows a parameter-free wrapper to impose the property around arbitrary architectures, including transformers with attention and LayerNorm. In a single-noise mismatch diagnostic for blind image denoising, WNE improved robustness of both CNNs and transformers with no measurable GPU overhead, whereas prior architectural NE baselines incurred up to 1.6x slowdown. The work characterizes the full NE function class and turns enforcement from an internal constraint into an input-output parameterization problem.
Key facts
- Normalization Equivariance (NE) is equivariance to global contrast and brightness transforms.
- Existing methods constrain internal layers to NE-compatible families, limiting compatibility with attention and LayerNorm.
- WNE is a parameter-free wrapper that enforces NE around any backbone.
- The method is based on the factorization: normalize-process-denormalize.
- WNE improves CNN and transformer robustness in blind denoising.
- No measurable GPU overhead with WNE.
- Architectural NE baselines incur up to 1.6x slowdown.
- The approach turns NE enforcement into an input-output parameterization problem.
Entities
Institutions
- arXiv