A Shadow Imaging Bilinear Model and Three-Branch Residual Network for Shadow Removal

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Current shadow removal methods often use shadow masks, which struggle with penumbras and small shadows, leading to complex pipelines.
• Existing techniques that only increase shadow brightness can introduce artifacts.

Purpose of the Study:

• To develop a more efficient and accurate single-image shadow removal method.
• To address limitations of existing shadow removal pipelines, including artifact generation and handling of complex shadow types.

Main Methods:

• Proposed a shadow imaging bilinear model to understand the shadow removal process.
• Designed a novel three-branch residual (TBR) network for shadow removal.
• Developed a single-stage network integrating illumination compensation, shadow reconstruction, shadow matte estimation, and shadow removal.

Main Results:

• The TBR network significantly shortens the shadow removal pipeline by eliminating the need for separate detection and refinement networks.
• The proposed method effectively restores light intensity in shadow areas while preserving non-shadow regions.
• Experimental results show superior performance compared to state-of-the-art shadow removal techniques.

Conclusions:

• The novel shadow imaging bilinear model provides insights into shadow removal complexities.
• The developed TBR network offers an efficient and effective solution for single-image shadow removal, outperforming existing methods.