Knowledge-Driven Deep Unrolling for Robust Image Layer Separation.

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Single-image layer separation is crucial for various vision tasks but is ill-posed.
• Existing methods struggle with complex regularizations and poor data fitting.
• A universal framework faces challenges in sharing visual cues across different tasks.

Purpose of the Study:

• To develop a generic optimization unrolling technique for adaptive image layer separation.
• To create a flexible and data-dependent framework by incorporating deep architectures into iterations.
• To address the limitations of existing methods in terms of complexity, adaptability, and task universality.

Main Methods:

• Proposed a general energy model with implicit priors based on maximum a posterior.
• Employed the alternating direction method of multipliers (ADMM) for the iteration mechanism.
• Utilized optimization unrolling with a general residual architecture prior and a task-specific prior.

Main Results:

• Achieved a straightforward, flexible, and data-dependent image separation framework.
• Successfully applied the method to four distinct tasks: rain streak removal, tone mapping, low-light enhancement, and reflection removal.
• Demonstrated significant qualitative and quantitative improvements over state-of-the-art methods.

Conclusions:

• The proposed method provides a versatile solution for multiple image separation tasks.
• The optimization unrolling technique effectively integrates deep learning into iterative separation processes.
• The framework shows superior performance and adaptability compared to existing approaches.