Contrastive Unfolding Deraining Network.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Image quality degradation from rain impacts outdoor vision systems.
• Single Image Deraining (SID) is crucial for enhancing visual data.
• Existing methods face challenges in accurately identifying and removing rain streaks.

Purpose of the Study:

• To propose a novel network, CUDEN, for effective single image deraining.
• To integrate traditional iterative algorithms with deep learning for improved performance and interpretability.
• To address the limitations of current deraining techniques by focusing on feature discovery.

Main Methods:

• Developed the Contrastive Unfolding DEraining Network (CUDEN).
• Introduced a dynamic multidomain translation (DMT) module for efficient mapping pair acquisition.
• Proposed a serial multireceptive field fusion (SMF) block for enhanced rain feature extraction.
• Pioneered the use of contrastive learning with perceptual loss (CPL) for the SID task.

Main Results:

• CUDEN demonstrates superior performance compared to state-of-the-art deraining networks.
• The proposed DMT and SMF modules effectively improve feature extraction and mapping.
• Contrastive perceptual loss aids in identifying optimal training directions.
• Experiments on synthetic and real-world data validate CUDEN's effectiveness.

Conclusions:

• CUDEN offers a powerful and interpretable solution for single image deraining.
• The integration of contrastive learning and deep unfolding networks advances the SID field.
• The novel modules contribute to robust and accurate rain removal from images.