ARF: Arbitrary Routing Framework for All-in-One Image Restoration.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Conventional image restoration requires separate models for each degradation type.
• Current all-in-one methods use a single model for all tasks, leading to inefficient resource allocation due to varying task complexities.

Purpose of the Study:

• To introduce the arbitrary routing framework (ARF) for efficient all-in-one image restoration.
• To dynamically adjust network structures based on image restoration task complexity.
• To improve both performance and computational efficiency in image restoration.

Main Methods:

• The arbitrary routing framework (ARF) integrates an arbitrary routing backbone (ARB) and task-specific neural architecture search (T-NAS).
• The ARB uses routing layers to enable flexible subnetwork configurations with minimal parameter overhead.
• T-NAS employs an efficiency-aware reward function to identify optimal subnetworks for specific restoration tasks.

Main Results:

• The ARF achieved a 0.31 increase in reconstruction PSNR across various image restoration tasks.
• Demonstrated a significant reduction in computational demands by 37.1% compared to the AirNet benchmark.
• The framework effectively allocates computational resources based on task-specific complexities.

Conclusions:

• The ARF offers a novel and efficient approach to all-in-one image restoration.
• Dynamic network adaptation based on task complexity leads to superior performance and reduced computational cost.
• This framework can be integrated with existing models to enhance their efficiency and effectiveness.