User-Guided Deep Human Image Matting Using Arbitrary Trimaps.

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Image matting is crucial for accurate foreground extraction but often requires precise trimaps.
• Existing deep learning models struggle with segmentation errors, leading to imperfect mattes.
• Current methods lack efficient user interaction for ambiguous situations.

Purpose of the Study:

• To develop a user-guided approach for practical human matting.
• To improve the efficiency and accuracy of foreground extraction.
• To reduce the user workload associated with trimap creation.

Main Methods:

• An end-to-end CNN architecture combining segmentation and matting stages.
• A residual-learning module for stroke-based user interaction.
• A novel trimap generation strategy and a large dataset of 12K human images.

Main Results:

• The proposed model efficiently corrects segmentation errors and supports arbitrary trimap inputs.
• Achieved superior performance over state-of-the-art automatic methods.
• Demonstrated competitive accuracy with high-quality trimaps.
• Outperformed separate trimap and alpha matte estimation models.

Conclusions:

• The user-guided interactive matting strategy significantly enhances practical human matting.
• The model offers flexibility in trimap input and foreground estimation.
• The approach provides a more efficient and accurate solution for foreground extraction.