Orthogonal Decoupling Contrastive Regularization: Toward Uncorrelated Feature Decoupling for Unpaired Image...

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Unpaired image restoration (UIR) is challenging due to the difficulty of obtaining perfectly matched degraded and clear image pairs.
• Existing methods often struggle with the inherent complexities of separating degradation effects from intrinsic image content.

Purpose of the Study:

• To develop a novel UIR method that effectively restores images without requiring paired data.
• To decouple degradation-related and degradation-unrelated features within images for improved restoration accuracy.

Main Methods:

• Introduced a Feature Orthogonalization Module optimized on the Stiefel manifold to ensure feature uncorrelation.
• Proposed a task-driven Depth-wise Feature Classifier to weight features based on degradation prediction relevance.
• Developed a robust training strategy using degradation-related proxies to mitigate reliance on single clear image pairs.
• Implemented a weighted PatchNCE loss to align specific feature types between restored and reference images.

Main Results:

• The proposed method successfully restores images by aligning degradation-related features with clear image characteristics and degradation-unrelated features with the input degraded image.
• The Feature Orthogonalization Module effectively decouples image features, enhancing the model's ability to distinguish between degradation and content.
• The robust training approach improves model generalization and performance across diverse degradation levels.

Conclusions:

• The novel UIR method offers a significant advancement in restoring images from unpaired data.
• The proposed feature decoupling and alignment strategy provides a robust framework for tackling UIR challenges.
• This work contributes a more reliable and effective approach to unpaired image restoration in computer vision.