Intrinsic Decomposition With Robustly Separating and Restoring Colored Illumination.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Intrinsic decomposition separates images into reflectance and shading, crucial for applications like image editing and augmented reality.
• Existing methods struggle with accurately separating colored illumination from reflectance and restoring it as shading.

Purpose of the Study:

• To propose a deep intrinsic decomposition method that effectively addresses the challenge of colored illumination separation.
• To improve the accuracy and robustness of intrinsic decomposition for enhanced image manipulation.

Main Methods:

• Developed a novel macro intrinsic decomposition network framework by transforming the process into intensity and chromaticity domains.
• Integrated multiple attention mechanism modules to enhance feature extraction.
• Proposed a skip connection module guided by deep features to filter physically irrelevant components.

Main Results:

• The proposed method outperforms state-of-the-art techniques across multiple datasets.
• Demonstrated robust separation of illumination from reflectance and accurate restoration into shading.
• Achieved visually superior image editing effects and enabled manipulation of scene lighting.

Conclusions:

• The novel deep intrinsic decomposition method offers significant improvements in separating and restoring image components.
• The framework provides finer intrinsic components through enhanced feature propagation and sub-constraints guidance.
• This advancement enables more sophisticated image editing and lighting manipulation in computer vision applications.