提出无需修改结构的归一化等变方法,提升图像去噪鲁棒性。
Normalization Equivariance for Arbitrary Backbones, with Application to Image Denoising
- 用包裹式设计实现归一化等变,兼容任意主干网络
- 在噪声水平不匹配时显著提升去噪性能,无额外计算开销
- 适合需要高鲁棒性的图像重建任务,如真实场景去噪
归一化等变(NE)是一种结构先验,可提升图像到图像任务中对分布偏移的鲁棒性。函数 $f$ 满足归一化等变性当且仅当 $f(a y + boldsymbol{1}) = a f(y) + boldsymbol{1}$,对所有 $a>0$ 且 $b ooldsymbol{1}$。现有方法要求每个内部层都使用符合 NE 约束的操作,增加运行时开销,并排除 softmax 注意力和 LayerNorm 等标准 Transformer 组件。本文提出包裹式归一化等变(WNE),一种无需参数的封装方法:先归一化输入,应用任意主干网络,再反归一化输出。我们证明所有 NE 函数均可通过此分解实现,因此该包装器能精确表示全部 NE 函数。在盲去噪任务中,将 CNN 与 Transformer 架构包裹后,在噪声水平不匹配情况下仍保持鲁棒性,且无明显 GPU 负担;而传统架构式 NE 方法最慢达 1.6 倍。
原文摘要 · Abstract (English)
Normalization Equivariance (NE) is a structural prior that improves robustness to distribution shift in image-to-image tasks. A function $f$ is normalization equivariant iff $f(a y + b\mathbf{1}) = a f(y) + b\mathbf{1}$ for all $a>0$ and $b\in\mathbb{R}$. Existing NE methods constrain every internal layer to NE-compatible operations. These constraints add runtime cost and exclude standard transformer components such as softmax attention and LayerNorm. We introduce Wrapped Normalization Equivariance (WNE), a parameter-free wrapper that normalizes the input, applies any backbone, and denormalizes the output. We prove every NE function admits this factorization, so the wrapper exactly parameterizes the class of NE functions. On blind denoising, wrapping CNN and transformer architectures improves robustness under noise-level mismatch with no measurable GPU overhead, while architectural NE baselines are up to $1.6\times$ slower.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品;中文卡片由大模型生成,请以原文为准。