Multiple-Step Model Training for Face Recognition

Recently, computer vision based on deep learning is developing rapidly. As an important branch in this area, face recognition has made great progress. The state of art has achieved 99.77% [1] pair-wise verification accuracy on LFW dataset. But the face dataset in the real application environment such as security checking in the station and bank account opening is much more complex than LFW because of face shelter, postures, uneven illumination and the different resolutions and so on. Except that, LFW dataset only contains the faces like western people but little of other area. Since faces from different areas have not consistent distribution, their methods always cannot achieve high recognition accuracy in practice. In this paper, aiming at Asian face, we propose a multiple-step model training method based on CNN network for real scene face recognition in the absence of large amounts of appropriate data. In the whole training process, each step plays an important role. For step1, it mainly enhanced the generalization ability of model by using a large-scale data set from different source. For step2, it improved the specificity of the model by using a smaller dataset which has closer data distribution in the real scene. And for the final step, metric learning is used to make the model more discriminative and expressive. Meanwhile, some strategy including data cleaning, data augmented and data balance are used in our method to improve the whole performance. Experiments show that this method can achieve high-performance for face recognition in the real application scene.