Improve-Center - A Deep Learning Face Representation Approach.

Convolutional neural networks have been widely used in object recognition, an important aspect of computer vision. The particular task of face recognition usually combines a softmax-loss function with some other loss function as the cost function in the training phase. In order to enhance the power of feature representation and speed up the training phase, this paper proposes a new supervised method called Improve-Center which is based on feature centers, the same as center-loss. It learns a center vector of features for every label and takes the feature of every sample closest to its center. This approach focuses on moving outer-space features closer to their center. The experimentation demonstrates that the approach is efficient. With softmax-loss and Improve-Center's joint supervision, a better model can be trained to make intra-class features more compact, and inter-class ones more discrete. In addition, the training process is faster.