Exploring the Efficiency of Face Recognition Learning Models Based on Specialization, Decomposition and Holistic Processing