Memory Uncertainty Learning for Real-World Single Image Deraining.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Deep neural networks have advanced single image deraining using synthetic data.
• A significant challenge remains in applying these methods to real-world images due to data discrepancies.

Purpose of the Study:

• To develop a semi-supervised method for effective real-world image deraining.
• To address the domain gap between synthetic and authentic rain images.

Main Methods:

• Proposing a stochastic memory network with self-supervised memory updates for capturing rain patterns.
• Implementing an uncertainty-aware self-training mechanism to transfer knowledge from supervised to unsupervised deraining.
• Creating a large-scale dataset of 10.2k real rain images to enhance diversity.

Main Results:

• The stochastic memory network effectively learns rain properties from both synthetic and real data.
• Uncertainty estimation guides knowledge transfer, improving deraining on unlabeled real-world images.
• The new dataset significantly expands the diversity of real rain scenes for training and evaluation.

Conclusions:

• The proposed memory-uncertainty guided semi-supervised method demonstrates superior performance in real-world image deraining.
• This approach effectively mitigates the synthetic-to-real domain gap.
• The developed dataset and method advance the field of image deraining.