Abstract
To address the problem of automatically detecting and removing the mask without user interaction, we present a GAN-based automatic approach for face de-occlusion, called Automatic Mask Generation Network for Face De-occlusion Using Stacked Generative Adversarial Networks (AFD-StackGAN). In this approach, we decompose the problem into two primary stages (i.e., Stage-I Network and Stage-II Network) and employ a separate GAN in both stages. Stage-I Network (Binary Mask Generation Network) automatically creates a binary mask for the masked region in the input images (occluded images). Then, Stage-II Network (Face De-occlusion Network) removes the mask object and synthesizes the damaged region with fine details while retaining the restored face's appearance and structural consistency. Furthermore, we create a paired synthetic face-occluded dataset using the publicly available CelebA face images to train the proposed model. AFD-StackGAN is evaluated using real-world test images gathered from the Internet. Our extensive experimental results confirm the robustness and efficiency of the proposed model in removing complex mask objects from facial images compared to the previous image manipulation approaches. Additionally, we provide ablation studies for performance comparison between the user-defined mask and auto-defined mask and demonstrate the benefits of refiner networks in the generation process.