GANMasker: A Two-Stage Generative Adversarial Network for High-Quality Face Mask Removal.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Deep learning excels at image inpainting, but removing large, complex face masks remains challenging.
• The COVID-19 pandemic increased interest in face mask removal technology.
• Limited paired masked/unmasked face datasets hinder development.

Purpose of the Study:

• To develop a robust deep learning model for accurate face mask removal.
• To address the challenge of reconstructing intricate facial features hidden by masks.
• To overcome the scarcity of paired masked and unmasked face image datasets.

Main Methods:

• A two-stage network combining an autoencoder for mask segmentation and a GAN with attention and Masked-Unmasked Region Fusion (MURF) for reconstruction.
• Training on the CelebA dataset, a large public collection of paired masked and unmasked faces.
• Evaluation using multi-scale masked faces to assess performance across various mask sizes and complexities.

Main Results:

• The proposed method significantly improves face mask removal compared to state-of-the-art techniques.
• Achieved a Peak Signal-to-Noise Ratio (PSNR) of 30.96 dB, an improvement of 4.18 dB over the second-best method.
• Demonstrated a 1% increase in Structural Similarity Index Measure (SSIM), reaching 0.95, indicating high visual fidelity.

Conclusions:

• The novel two-stage network effectively removes face masks, generating realistic and accurate unmasked faces.
• The attention and MURF mechanisms in the GAN enhance the focus on masked regions for superior reconstruction.
• The method offers a significant advancement in unmasking faces, particularly for complex and large masks.