We found a match
Your institution may have access to this item. Find your institution then sign in to continue.
- Title
DIQA-FF:dual image quality assessment for face frontalization.
- Authors
Duan, Xinyi; Liu, Hao; Liang, Jiuzhen
- Abstract
Face frontalization is a process of synthesizing a realistic and identity-preserving face view from different face pose images. It is an essential preprocessing step for face recognition. As Generative Adversarial Networks (GANs) are applied to face frontalization tasks, the synthesized high-quality frontal faces are required to be more realistic and highly recognizable. To increase the authenticity and recognizability of generated frontal face images, we propose a face frontalization algorithm with dual image quality assessment (DIQA-FF) in this study. This approach assesses the generated images' quality from the discriminator and the generator. We use the Gaussian mixture model (GMM) in the discriminator to give the generated images a probability distribution that is more similar to the authentic images, increasing the images' authenticity. To improve the recognizability of the images, we caculate the similarity of the authentic images and the generated images and maximize the resemblance between them in the generator. The results of the experiments demonstrate that our method can transform the side face image into a complete and accurate front face image. Additionally, the generated frontal faces perform better for face recognition. Quantitative and qualitative evaluations on both controlled and in-the-wild databases show that our frontalization approach not only outperforms the state of-the-art in terms of authenticity but also in terms of recognizability.
- Publication
Multimedia Tools & Applications, 2023, Vol 82, Issue 25, p39503
- ISSN
1380-7501
- Publication type
Article
- DOI
10.1007/s11042-023-15084-8