A comparison of classification approaches for deep face recognition

Face recognition is a mature domain with lots of different techniques proposed in the literature. Convolutional neural networks have been the most successful approach to face recognition problem recently. In this work, performance of three different face recognition models are compared. Features are extracted using a pre-trained convolutional neural network. The first model is trained using the available face images of a subject as positive samples and all other available face images as negative samples. In this case, recognition is done by looking at the scores of binary verification models. The second model is trained using the available face images of a subject as positive samples and all other registered subjects' available face images as negative samples (one versus all). The third model is trained using the training samples of a subject and other subjects one by one (one-versus-one). These models are compared in terms of recognition accuracy, efficiency, suitability to incremental learning, complexity, rejection accuracy, score normalization necessity. We show that even when other enrolled subjects are not taken into account as negative samples, verification model is easy to manage and outperforms other models.