Complementary Mixture-of-Experts and Complementary Cross-Attention for Single Image Reflection Separation in the Wild

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Single Image Reflection Separation (SIRS) is crucial for reconstructing clear images from those with superimposed reflections.
• Existing learning-based SIRS methods excel on standard images but falter with complex, real-world reflections.

Purpose of the Study:

• To develop a universal SIRS framework capable of handling diverse and challenging reflection artifacts.
• To improve the robustness and accuracy of reflection separation in unconstrained environments.

Main Methods:

• A flexible dual-stream architecture incorporating a Mixture-of-Experts mechanism for adaptive patch-based expert assignment.
• Utilizing multi-head attention from Transformers to capture and leverage both high and low cross-correlations between image streams.
• Implementing adaptive inter-stream feature interactions for enhanced separation.

Main Results:

• The proposed universal SIRS framework demonstrates significant improvements over state-of-the-art methods.
• Qualitative and quantitative evaluations on diverse real-world datasets confirm the method's effectiveness.
• The Mixture-of-Experts and Transformer-based attention mechanisms contribute to superior reflection separation.

Conclusions:

• The developed dual-stream framework offers a robust solution for Single Image Reflection Separation.
• The adaptive feature interaction and expert assignment mechanisms are key to handling complex reflections.
• This work advances the capabilities of reflection removal in challenging, real-world image scenarios.