Mask removal : Face inpainting via attributes

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• The COVID-19 pandemic necessitated mask-wearing, leading to facial occlusion challenges.
• Existing face inpainting methods struggle to achieve high-fidelity results for mask removal.

Purpose of the Study:

• To develop an advanced face inpainting network for high-fidelity mask removal.
• To leverage known facial attributes for more accurate face reconstruction.

Main Methods:

• A novel dual pipeline Generative Adversarial Network (GAN) based approach was proposed.
• A reconstructive path utilized ground truth for prior distribution, while a generative path predicted masked regions.
• A synthetic facial occlusion was created to simulate real mask removal scenarios.

Main Results:

• The proposed method generated faces more closely aligned with real attributes (e.g., nose, gender, makeup).
• The network ensured semantic and structural rationality in the inpainted regions.
• Experimental results demonstrated superior performance compared to state-of-the-art methods.

Conclusions:

• The novel face inpainting network effectively removes masks while preserving facial attribute accuracy.
• This approach offers a significant improvement in generating realistic and semantically coherent facial images post-mask removal.