Neural Architecture Search for Mobile Face Recognition

Biometrics is a rapidly growing technology that aims to identify or verify people's identities based on their physical or behavioral properties. With the rapid growth of smartphone users, the interest in secure authentication using biometric technology to authorize and identify the application user has been increased. With its high social acceptability, face recognition is one of the most convenient and accurate biometric recognition technologies which is used more and more in mobile and embedded systems for unlocking, application login and mobile payment. To enable face recognition on low computational powered devices, the model needs to be accurate, small and fast. With the rise of deep learning technology, face recognition systems were able to achieve notable verification performances. However, most high-accurate face recognition models rely on very deep Convolutional Neural Networks (CNNs) and therefore, require a high amount of computational resources, which makes them unfeasible for mobile and embedded systems. Recently, a great progess has been made in designing efficient face recognition systems by utilizing lightweight deep learning network architectures designed for common computer vision tasks for face recognition. However, none of these works designed a network specifically for the face recognition task, rather than adopting existing architectures designed for common computer vision tasks. With the development of AutoML, Neural Architecture Search (NAS) has shown excellent performance in many computer vision tasks and was able to automatically design highly efficient network architectures that outperform existing manually designed architectures. This thesis utilizes Neural Architecture Search (NAS) to automate the process of designing highly efficient neural architectures for face recognition. While other works focused on manually designing efficient architectures, this process has not been made automatic for the face recognition setting. This is the first work that utilizes Differentiable Architecture Search (DARTS) for face recognition and introduces a new search space. Based on DARTS, this thesis introduces a new network architecture named DartFaceNet. Evaluation on a variety of large-scale databases proves that DartFaceNet is able to achieve high performance on major face recognition benchmarks. Using DartFaceNet architecture and different embedding sizes, this thesis introduces three face recognition models with less than 2 million parameters. Trained with ArcFace loss on MS1MV2 dataset, DartFaceNet-256 achieves 99.5% on LFW with only 0.991 million parameters and 587.11 FLOPs which is comparable to the efficient light-weight architecture from MobileFaceNets. With less than 2 million parameters, DartFaceNet-512 outperforms existing light-weight models under 5 million parameters on CA-LFW with 95.333%. Also this thesis provides evaluation on the large-scale databases IJB-B, IJB-C and the MegaFace challenge. With only 0.925 million parameters and a memory footprint of 3.7 Megabytes, DartFaceNet-128 achieves the best trade-off between performance and model size among all DartFaceNet models