Learning deep discriminative features based on cosine loss function

Abstract

Deep feature representation is widely used in various visual applications. However, the feature extracted by the convolutional neural networks (CNNs) is inappropriate for cosine similarity measurement. Because the classical CNNs are designed for classification rather than for similarity comparison. A novel cosine loss function for learning deep discriminative features, which are fit to the cosine similarity measurement, is designed. The loss can constrain the distribution of the features in the same class to be in a narrow angle region. Furthermore, a discriminative feature learning network framework and its corresponding two-stage learning method to learn the parameters is proposed. Experimental results show that the proposed method achieves state-of-the-art performance on the public Cifar10 and Market1501 datasets.