Self-Supervised Matting-Specific Portrait Enhancement and Generation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Alpha matting is an ill-posed problem in image processing.
• Existing matting models struggle with complex portrait images.
• Manual annotation of alpha mattes is time-consuming and expensive.

Purpose of the Study:

• To propose a novel approach for improving alpha matting by enhancing input images.
• To leverage Generative Adversarial Networks (GANs) for semantic image transformations.
• To reduce the dependency on expensive manual annotations for training matting models.

Main Methods:

• Inverting input portraits into the latent space of StyleGAN.
• Optimizing multi-scale latent vectors using tailored losses for matting-specificity and subtle modifications.
• Utilizing the generative capabilities of StyleGAN to create pseudo ground truth (GT) data.

Main Results:

• The proposed method refines real portrait images, significantly boosting the performance of arbitrary matting models.
• Enhanced portrait images lead to a large margin improvement in automatic alpha matting.
• Generated pseudo GT data effectively addresses the annotation cost issue.

Conclusions:

• Exploring GAN latent spaces offers a new perspective for solving the alpha matting problem.
• The method enhances image compatibility with existing matting models, improving their accuracy.
• This approach provides a cost-effective solution for data augmentation in alpha matting research.