Nalini K Ratha - Deep Learning-Based Face Analytics

More information about this series at http://www.springer.com/series/4205

This Springer imprint is published by the registered company Springer Nature Switzerland AG

The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland

Automatic face recognition is the problem of identifying a person from an image or a video. Due to the ubiquity of cameras and prevalence of social media networks, automatic face recognition has applications in access control, homeland security, rescuing exploited children, HCI interfaces, etc. Recent years have seen significant progress in automatic face recognition technology, largely due to improvements in deep convolutional network designs and the availability of large datasets []. In this article, we summarize recent works in automatic face recognition , focusing on methods using deep convolutional neural networks (CNNs).

The problem of face recognition can be divided into face identification and face verification . The standard approach for training a CNN for solving these problems includes four steps: face detection, alignment, representation, and classification (Fig. ). Identification is the problem of assigning an identity to an image from a list of identities. From another perspective, this can be considered as trying to retrieve the best matching face from a gallery for a given probe image. On the other hand, face verification involves verifying whether two face images are of the same person. This is usually performed by computing the similarity between feature representations of the two faces. Both identification and verification have benefited immensely from developments in deep learning algorithms and more advanced CNN architectures.

In addition to improved architectures, face recognition has seen significant progress in the design of effective loss functions for training CNNs. Both face identification and verification aim to learn representations which have low intra-class variations and high inter-class variations. Several loss functions have been proposed over the past few years which encourage representations with these properties. Most of these [] modify the common softmax loss using additional constraints on the features which lead to compact and discriminative representations of faces and thus, performance improvements in face recognition .