Unsupervised Deraining: Where Asymmetric Contrastive Learning Meets Self-Similarity

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Supervised deraining methods struggle with real-world data due to domain gaps.
• Existing methods often treat image and rain layers independently, ignoring their relationship.

Purpose of the Study:

• To develop an unsupervised deraining method that generalizes better to real-world rainy scenes.
• To address the limitations of synthetic training data and independent layer processing in current deraining techniques.

Main Methods:

• Proposed an unsupervised non-local contrastive learning (NLCL) deraining method.
• Exploited intra-layer similarity and inter-layer dissimilarity using non-local self-similarity patches.
• Introduced an asymmetric contrastive loss based on the dimensional discrepancy between image and rain patches.
• Collected a large-scale, high-resolution real-world rainy dataset.

Main Results:

• The NLCL method effectively differentiates rain from clean images by learning compact representations and discriminative decompositions.
• The asymmetric loss enhances decomposition by modeling compactness discrepancies.
• Demonstrated state-of-the-art performance on various real-world rainy datasets.

Conclusions:

• Unsupervised contrastive learning effectively bridges the domain gap in deraining.
• Leveraging both intra-layer similarity and inter-layer dissimilarity is crucial for robust deraining.
• The proposed method offers a significant advancement for real-world image deraining applications.