Reproduce “Semantic-Guided Zero-Shot Learning for Low-Light Image/Video Enhancement” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Shen Zheng, Gaurav Gupta
The paper introduces SGZ, a semantic-guided zero-shot learning framework for low-light image/video enhancement that integrates high-level semantic information without requiring paired or unpaired training data. This lightweight model achieves real-time performance, significantly improving perceptual quality and boosting accuracy for downstream tasks like object detection and semantic segmentation, particularly in extreme low-light conditions.
Key insights
- Integrating high-level semantic information into zero-shot learning for image enhancement can significantly improve visual quality and the performance of downstream computer vision tasks.
- A pre-trained and frozen semantic segmentation network can effectively provide "unsupervised" semantic guidance for image enhancement by leveraging its feature outputs in a non-reference loss function, thus avoiding burdensome annotation requirements.
- Combining a lightweight enhancement factor prediction network with a recurrent enhancement module enables extremely fast, real-time processing and a minimal model footprint suitable for video enhancement and deployment on resource-constrained devices.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.