A Lightweight Attention Model for Face Recognition

Facial recognition is one of the most popular biometric methods in the real world. To address this task, various methods have been proposed especially deep learning-based approaches. However, most of them usually focus on improving performance by building deeper and more complex networks. This makes an important limitation for the ability to deploy them on embedded or mobile devices without GPU. In this paper, we introduce an efficient network called SeesawAttentionFaceNet which is a hybrid of SeesawFaceNet and CBAM attention to improve the performance of face recognition while keeping the simple to suitable for edge devices. The experiments conducted on various datasets have shown that our proposed framework outperforms the other state-of-the-art lightweight models for facial recognition. Moreover, we also provide an ablation study to demonstrate that when the face area of the removed position increases, the recognition results typically are reduced. This has important implications in feature extraction and processing of facial images.