Aligning Network Equivariance with Data Symmetry: A Theoretical Framework and Adaptive Approach for Image Restoration

Researchers propose a theory for dataset-level non-strict symmetry and show how it connects to restoration equivariance and error bounds.

The framework suggests that aligning model symmetry with real-world data symmetry can lower expected risk in image restoration.

They also introduce a sample-adaptive equivariant network using a hypernetwork to adjust equivariant convolutions per input.

The method outperforms standard models and earlier equivariant baselines on denoising, super-resolution, and deraining benchmarks.