Joint face normalization and representation learning for face recognition

Identity-independent factors, such as variations of pose, expression, illumination, etc., are the key challenges in face recognition. To avoid the effects of these factors, existing face recognition methods usually adopt two approaches: pose-invariant face feature extracting and face normalization before feature extraction. Contrary to these, we propose a single deep model jointly performing face normalization and representation learning tasks for face recognition, named normalization and reconstruction general adversarial network (NRGAN). First, the unified NRGAN model can boost the performance of the two tasks for each other. Second, NRGAN can synthesize normalized face images without the requirement of paired data, which makes our method have better generalization ability to the uncontrolled environment. Third, a factor-invariant identity disentanglement training strategy is introduced to decouple the identity feature representation from other factors without using any of these factors' labels. Extensive experiment results on four currently popular face datasets demonstrate the effectiveness of NRGAN on both normalized face synthesis and face recognition tasks.