End-to-end Facial Recognition Deep Learning Model Specialized for Facial Angle using Gray Image

Face-recognition using deep-learning techniques typically face problems with learning about angles and with flexibility. In this study, we propose a face recognition system specialized for angles by using gray images. The K-face dataset was used to perform learning, along with triplet loss to achieve angle-specific learning. We reconstructed the model structure of Inceptionresnetv2, a model with excellent face recognition performance and the proposed model achieved an accuracy of approximately 96.9 %. In addition, the performance verification for each illuminance and angle was conducted and showed that the proposed model was flexible to the angle and illuminance. To confirm the applicability of model in the real environment, 21 subjects were recruited and verified. Transfer learning was performed to ensure the flexibility of the model in the real environment and gray images were used to reduce the effect of illuminance. Consequently, we obtained approximately 81% of the results and it was demonstrated that the proposed face recognition system can be applied in a real environment.