From Controlled Scenarios to the Real World: Cross-Domain Degradation Pattern Matching for All-in-One Image...

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• All-in-One Image Restoration (AiOIR) aims to restore images degraded by multiple patterns using a single model.
• Existing methods struggle with sample-wise supervision, entangling degradation with content and suffering from domain shift, limiting real-world generalization.

Purpose of the Study:

• To develop a Unified Domain-Adaptive Image Restoration (UDAIR) model.
• To decouple degradation features from image content.
• To improve model generalization to real-world, unseen degradation patterns.

Main Methods:

• Implemented a Cross-Sample Contrastive Learning mechanism with a codebook to learn discrete degradation prototypes.
• Utilized a correlation alignment-based test-time adaptation to bridge the domain gap during inference.
• Contrasted samples with shared degradations but diverse content to learn robust universal prototypes.

Main Results:

• Achieved new state-of-the-art performance on the AiOIR task across 10 datasets.
• Demonstrated effective degradation identification under multiple degradation patterns via feature clustering.
• Showcased robust generalization capabilities to real-world scenarios through qualitative comparisons.

Conclusions:

• The proposed UDAIR model successfully decouples degradation from content and bridges the domain shift.
• The technical modules contribute significantly to improved performance and generalization.
• UDAIR offers a robust solution for real-world image restoration challenges.