Paired regions for shadow detection and removal.

Area of Science:

• Computer Vision
• Image Processing
• Computational Photography

Background:

• Shadows in natural images pose challenges for image analysis and manipulation.
• Traditional methods often rely on pixel or edge information, limiting their effectiveness.

Purpose of the Study:

• To develop an effective region-based approach for shadow detection and removal from single natural images.
• To improve upon existing methods by incorporating relative illumination conditions between image regions.

Main Methods:

• A region-based approach is employed, analyzing segmented regions and their appearance.
• Pairwise classification predicts relative illumination conditions between regions.
• Graph-cut is utilized for shadow and non-shadow region labeling.
• Image matting refines detection, and a lighting model recovers shadow-free images.

Main Results:

• The method demonstrates effectiveness in shadow detection and removal.
• Evaluation on existing and newly created datasets validates the approach.
• Analysis of features for unary and pairwise classification provides insights.

Conclusions:

• The proposed region-based method offers a robust solution for shadow detection and removal.
• The integration of pairwise classification and graph-cut enhances accuracy.
• The developed dataset facilitates quantitative evaluation of shadow removal techniques.