Model Architecture
We propose a Semantic-Guided Zero-Shot (SGZ) network for low-light image enhancement. It consists of:
- An Enhancement Factor Extraction (EFE) network for estimating pixel-wise light deficiency.
- A Recurrent Image Enhancement (RIE) network to progressively enhance the low-light image.
- An Unsupervised Semantic Segmentation (USS) network for preserving the semantic information.
During training, both RIE and USS have frozen parameters and they output the loss to update EFE. During testing, EFE and RIE are used sequentially to enhance a low-light image.
