Facial Adversarial Sample Augmentation for Robust Low-Quality 3D Face Recognition

Compared with traditional 3D face recognition tasks using high precision 3D face scans, 3D face recognition based on low-quality data captured by consumer depth cameras is more practicable for real-world applications. However, it is also more challenging to deal with the variations of facial expressions, poses, occlusions, data noises, scanning distance, and so on. In this paper, we propose a novel robust low-quality 3D face recognition method based on Facial Adversarial Sample Augmentation, namely FASA-3DFR. It consists of two modules, namely facial adversarial sample generation, and facial adversarial sample training. For the first module, to enlarge the diversity of facial adversarial samples and boost the robustness of 3DFR, we propose to utilize the Kullback-Leibler divergence to maximize the distribution distance between the original and adversarial facial samples. For the second module, a distribution alignment loss is designed to make the distribution of facial adversarial samples gradually close to the one of the original facial samples, and the common and valuable information from both distributions can be effectively extracted. Extensive experiments conducted on the CAS-AIR-3D Face database show the effectiveness of the proposed method.