Adaptive Face Recognition Using Adversarial Information Network.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Face recognition models struggle with domain shift, where training and testing data differ due to factors like pose and skin tone.
• Unsupervised domain adaptation uses pseudo-labels from clustering, but often misses hard samples and creates intra-domain gaps.
• This gap between pseudo-labeled and unlabeled data hinders discrimination in face recognition.

Purpose of the Study:

• To propose a novel Adversarial Information Network (AIN) for robust face recognition under domain shift.
• To mitigate the intra-domain gap caused by pseudo-labeling in unsupervised domain adaptation.
• To improve cross-domain generalization for face recognition models.

Main Methods:

• Introduced a novel adversarial mutual information (MI) loss to adaptively modify target prototype positions, aiding clustering of unlabeled images.
• Employed a graph convolution network to predict linkage likelihoods and generate more reliable pseudo-labels.
• Evaluated the method on domain adaptation across poses, image conditions, and different skin tones.

Main Results:

• The proposed AIN effectively mitigates the intra-domain gap, leading to more discriminative features.
• AIN demonstrates significant improvements in cross-domain generalization for face recognition tasks.
• Achieved state-of-the-art performance on the RFW dataset.

Conclusions:

• The Adversarial Information Network (AIN) offers an effective solution for domain adaptation in face recognition.
• The combination of adversarial MI loss and graph convolution network enhances pseudo-label reliability and reduces domain gaps.
• AIN represents a significant advancement in achieving robust and accurate face recognition across diverse real-world conditions.