A Metric Learning Method Based on Damped Momentum with Threshold

Abstract

The convolutional neural networks in deep learning have become one of the mainstream algorithms of face recognition technology. Moreover, metric learning is also an important method to train deep learning models, as its ability of verification is very powerful, especially for the face images which are often used in CNNs. Recently, a new type method of metric learning named Center Loss has been proposed. It is simple to use and can enhance the model performance obviously. However, since the updating mechanism of Center Loss is simplistic, it can hardly process large-scale data when the categories are too much. This paper proposes an improved algorithm of Center Loss to accelerate the updating process of feature centers of original algorithm with a damped momentum, which urges deep learning models to have more rapid and steady convergence and better performance. Meanwhile, almost no additional computation cost is added since the new method has an optional threshold. The experimental results show that the improved Center Loss algorithm can further improve the recognition ability of the model, which is very helpful to enhancing the user experience of complex face recognition systems.