Unsupervised face Frontalization for pose-invariant face recognition

作者：

Highlights：

• In this paper, we proposed a novel unsupervised face frontalization method for pose-invariant face recognition. The highlights of this paper include:

• Propose an unsupervised Pose Conditional CycleGAN (PCCycleGAN) to generate photorealistic frontal face images from arbitrary poses.

• Introduce a conditional pose vector to control different pose generation of the inverse mapping of face frontalization.

• Introduce feature space preception loss and adopt pixel loss and identity preserving to synthesize frontal face image with high quality as well as maintain the identity information.

摘要

In this paper, we proposed a novel unsupervised face frontalization method for pose-invariant face recognition. The highlights of this paper include:•Propose an unsupervised Pose Conditional CycleGAN (PCCycleGAN) to generate photorealistic frontal face images from arbitrary poses.•Introduce a conditional pose vector to control different pose generation of the inverse mapping of face frontalization.•Introduce feature space preception loss and adopt pixel loss and identity preserving to synthesize frontal face image with high quality as well as maintain the identity information.

论文关键词：Face frontalization generative adversarial network pose-invariant face recognition

论文评审过程：Received 10 July 2020, Revised 31 October 2020, Accepted 10 December 2020, Available online 13 December 2020, Version of Record 28 December 2020.

论文官网地址：https://doi.org/10.1016/j.imavis.2020.104093