Pre-Trained Feature Fusion and Multidomain Identification Generative Adversarial Network for Face Frontalization

Abstract

The study of face frontalization is essential for improving face recognition accuracy in extreme pose scenarios. Mainstream methods like TP-GAN, CAPG-GAN, etc., have made meaningful contributions. However, they still suffer from two problems: the lack of extracted feature diversity and the blurred details in generated images. This paper proposes a pre-trained feature fusion and multi-domain identification generative adversarial network (PM-GAN) for face frontalization: the features of the model pre-trained on large-scale datasets are fused with the original features of the encoder to enhance the diversity and robustness of features. In order to fuse features more effectively, we design a novel feature fusion module (FFM). In addition, a group of global and local discriminators is introduced to reinforce the local details and realism of generated frontal faces. Experimental results show that our proposed method outperforms state-of-the-art methods on M2FPA and CAS-PEAL datasets.

Keywords

• Face frontalization
• transferring pre-trained network
• feature fusion
• detail optimization
• generative adversarial network