MWFormer: Multi-Weather Image Restoration Using Degradation-Aware Transformers.

Area of Science:

• Computer Vision
• Image Restoration
• Deep Learning

Background:

• Image restoration under adverse weather is crucial for computer vision.
• Existing methods struggle with multiple simultaneous weather degradations (e.g., rain and snow).

Purpose of the Study:

• To develop a unified architecture capable of restoring images from multiple weather conditions.
• To enhance adaptability and controllability in weather restoration tasks.

Main Methods:

• Proposed MWFormer, a vision Transformer utilizing hyper-networks and feature-wise linear modulation.
• Employed contrastive learning for distortion-aware feature embeddings to predict weather types.
• Enabled adaptive parameter modulation for multi-weather restoration without retraining.

Main Results:

• MWFormer achieved significant performance improvements on multi-weather restoration benchmarks.
• Demonstrated superior results compared to state-of-the-art methods with low computational cost.
• Showcased the versatility of hyper-networks for enhancing various architectures.

Conclusions:

• MWFormer offers a unified and adaptable solution for complex multi-weather image restoration.
• The proposed hyper-network approach enhances performance and controllability.
• This work advances the field of robust image restoration for real-world applications.