Face recognition based on general structure and angular face elements

Face recognition methods achieve their highest accuracy when faces are captured in the frontal view. However, the accuracy of these methods decreases when the angle of a person’s face changes relative to the camera. The problem of pose variation in face recognition can be addressed in either the feature space or the image space. While generating a frontal face in the image space can be costly, face recognition relies on feature vectors. This research proposes a method to effectively modify the feature vectors of angular face images. The proposed method involves segmenting the angular face image elements using a fine-tuned DeepLabv3 network. Subsequently, an Attention-equipped U-Net network transforms the angular face elements into frontal face elements. Features are then extracted from the normalized element image using a deep convolutional network to capture general information about the features. Detailed features are additionally obtained using the pre-trained VGGFace architecture. These detailed features are combined with the overall feature vector extracted from the facial elements, enabling the feature vector of an angular image to exhibit similarity to the feature vector of a frontal image of the same person. The proposed model was trained using the MUT1NY and FERET datasets and evaluated on a subset of the FERET dataset, which includes images from 200 individuals captured at various angles. The proposed model achieved an impressive average recognition accuracy of 99.81% on this evaluation set. © The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2024.