Purohit K, Suin M, Rajagopalan A N, et al. Spatially-adaptive image restoration using distortion-guided networks[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision. 2021: 2309-2319.
实验效果图:
Rain-drop Removal
Shadow Removal
解决什么问题?
SPAIR结构
(1)A localization network that identifies degraded pixels.
(2) A restoration network that exploits knowledge from the localization network in filter and feature domain to selectively and adaptively restore degraded pixels
The architecture of SPAIR.
核心模块
1. Spatial Feature Modulator (SFM)
- Efficacy: SFM is to perform distortion-guided spatial-varying feature normalization.
- Process: Given the fused features and the predicted mask
, we calculate the modulated features
.
- Studies(Justin Johnson et al.2016) show the correspondence between
- Feature mean → global semantic information.
- Feature variance → local texture.
- SFM modulates features at M = 1 to match the feature statistics of features at M = 0.
2 Mask-guided Sparse Convolution (SC)
- Efficacy: Facilitates selective restoration of highly degraded regions, and simplifies the learning process.
- Process: Given the fused features and the predicted mask
, we calculate the sparse features
.
Actually, means that->