| Abstract |
Face inpainting is a currently developing technology with multiple real-life applications. The main objective of this thesis is to enhance an intelligent face inpainting system, which helps in raising the efficiency of face recognition systems. The enhanced system uses a Generative Adversarial Network (GAN) to inpaint face images. The network consists of a subnet to predict landmarks, and another one to generate a new pixel for missing parts based on the predicted landmarks. To reconstruct an intelligent system capable of face inpainting correctly, the system was trained using two databases, the Large-scale CelebFaces Attributes Dataset (CelebA) beside Novel Landmarked Face Database for Arab Celebrities, and evaluated them. The face inpainting system approach consists of four steps as follows:
1. Create a new Arab face database.
2. Construct our enhanced face inpainting model.
3. Train the model on two datasets different in ethnicity to enhance face landmark guidance and face completion.
4. Inpainting using landmark guided.
Finally, an intelligent system is developed which can inpaint a complete face image correctly. From the quantitative results, the proposed method achieves the maximum score of 34.97, 0.989, and 1.82 on PSNR (Peak Signal to Noise Ratio), SSIM (Structure Similarity Index Measure), and FID (Fréchet Inception Distance) metrics, respectively. This approach is implemented by using Python programming language.
|
