Deep learning has been widely used in solving the various challenges that are encountered in the field of face recognition. Pose variations, illumination changes, aging of the individual whose face is being recognized, and changes in the spectral distribution of the images of the faces all pose challenges for recognition systems. To combat these issues, various modalities of facial data have been incorporated into face recognition systems. The different data types that are used in deep learning for face recognition include RGB data, 3D facial data, video data, near-infrared (NIR) data, and synthetic data. RGB data is the most common modality used for face recognition, but other modalities are often incorporated into address specific recognition challenges. Commonly used deep learning models in face recognition include convolutional neural networks (CNNs), ResNets, and VGG models. Additionally, various innovations to these models have been developed, such as using different data augmentations specific to the task of face recognition, fusing features from different data types, creating new types of loss functions, such as ArcFace, and using attention mechanisms. Recognition models that use multiple data types and address spectral variations tend to experience improvements in recognition rates of between 0.5% and over 15%. Additionally, recent developments in face recognition that address issues such as data privacy, robustness against adversarial attacks, and fairness among different demographics highlight the shift towards holistic and ethical solutions for face recognition. The integration of diverse data types into deep learning models for face recognition and the use of diverse evaluation methodologies to test these models are crucial elements in overcoming the various challenges that are experienced within the field of deep learning and face recognition.
Anarbayeva et al. (Thu,) studied this question.