Triple-Level Model Inferred Collaborative Network Architecture for Video Deraining.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Video deraining is crucial for outdoor vision systems.
• Designing optimal architectures for diverse rain patterns remains challenging.
• Existing methods struggle with varied rain streak distributions.

Purpose of the Study:

• To develop a model-guided framework for automatic video deraining architecture search.
• To address limitations in handling various rain streak distributions.
• To improve deraining performance in terms of fidelity and temporal consistency.

Main Methods:

• Introduced a Triple-level Model Inferred Cooperating Searching (TMICS) framework.
• Designed a hyper-parameter optimization model for task variables and parameters.
• Developed a collaborative structure with Dominant Network Architecture (DNA) and Companionate Network Architecture (CNA) using Attention-based Averaging Scheme (AAS).
• Implemented a macroscopic structure search using Optical Flow Module (OFM) and Temporal Grouping Module (TGM).
• Utilized differentiable neural architecture search from a compact candidate set.

Main Results:

• The TMICS framework demonstrates significant improvements in fidelity and temporal consistency.
• The proposed collaborative structure effectively handles diverse rain streak distributions.
• Automatic architecture discovery leads to superior rain removal performance compared to state-of-the-art methods.

Conclusions:

• The TMICS framework offers an effective solution for video deraining.
• The model-guided approach and cooperating optimization enhance deraining capabilities.
• This work advances the field of computer vision for adverse weather conditions.