Cycle contrastive adversarial learning with structural consistency for unsupervised high-quality image deraining...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Unsupervised single image deraining (SID) methods struggle with semantic representation and content preservation.
• Existing methods often fail to completely separate rain layers from image content, limiting deraining quality.

Purpose of the Study:

• To develop a novel unsupervised SID framework that enhances deraining performance by focusing on semantic representation and content integrity.
• To introduce a new approach for high-quality image reconstruction and effective rain-layer stripping in unsupervised deraining.

Main Methods:

• A novel cycle contrastive adversarial framework incorporating cycle contrastive learning (CCL) and location contrastive learning (LCL).
• CCL pulls similar features and pushes dissimilar ones in latent spaces for reconstruction and rain-layer stripping.
• LCL constrains mutual information at the same location across different exemplars to preserve content.
• Integration of Segment Anything Model (SAM) for structural-consistency regularization.
• Introduction of Vision Transformer (ViT) and a multi-layer channel compression attention module (MCCAM) for improved feature representation.

Main Results:

• The proposed CCLformer demonstrates advantageous unsupervised image deraining performance.
• Extensive experiments validate the superiority of the CCLformer method.
• The effectiveness of individual modules within CCLformer is confirmed through rigorous testing.

Conclusions:

• CCLformer offers a significant advancement in unsupervised single image deraining.
• The combination of CCL, LCL, SAM, and ViT-based architecture effectively addresses limitations of previous methods.
• The framework successfully achieves high-quality deraining while preserving crucial image content.