Path-Restore: Learning Network Path Selection for Image Restoration.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Deep Convolutional Neural Networks (CNNs) excel in image restoration but face computational challenges.
• Image restoration complexity varies across regions due to differing distortions and content.

Purpose of the Study:

• To develop an efficient image restoration method that addresses the computational burden of deep CNNs.
• To dynamically adapt restoration strategies based on the difficulty of image regions.

Main Methods:

• Proposed Path-Restore, a multi-path CNN with a reinforcement learning-trained pathfinder.
• Introduced a difficulty-regulated reward incorporating performance, complexity, and region restoration difficulty.
• Investigated a policy mask for joint processing of all image regions.

Main Results:

• Achieved comparable or superior performance to existing methods with reduced computational cost.
• Demonstrated effectiveness in real-world denoising with spatially varying noise.
• Outperformed RIDNet by 2.7x on the Darmstadt Noise Dataset while maintaining performance.

Conclusions:

• Path-Restore offers an efficient and effective solution for image restoration tasks, particularly those with spatially varying degradations.
• The dynamic path selection mechanism significantly reduces computational requirements.
• The method shows promise for practical applications requiring high-quality image restoration with limited resources.