DIQA-FF:dual Image Quality Assessment for Face Frontalization

Xinyi Duan,Hao Liu,Jiuzhen Liang
DOI: https://doi.org/10.1007/s11042-023-15084-8
IF: 2.577
2023-01-01
Multimedia Tools and Applications
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.
What problem does this paper attempt to address?