Bio-identification by deep neural networks in such devices with limited computational power and memory capacity as mobile phones has become an essential but challenging task today. As faces have rather invariable features among human biometric features, face recognition is considered as the most important biometric identification task. Face recognition has been widely used for user authentication in security systems such as electronic payment systems as it can identify faces from facial images from photographs or videos, and it has been studied for years.
The face recognition system using convolutional neural networks is considered as the best method among the existing ones. Face recognition network models that have been developed recently and have proved to be superior in performance cannot be used for real-time face recognition in devices with limited computational resources such as low base computers or mobile phones because their structure is very complex and they need a large amount of computation. What is more, reducing the number of layers continuously to reduce computational burden affects recognition performance.
In previous studies, several methods to improve the trade-off between speed and recognition performance were proposed. One of them is GhostFaceNets which uses Ghost module to reduce the feature map redundancy, where the trade-off between speed and recognition performance is improved by extracting less repetitive feature maps with small amount of computation. In GhostFaceNets, they improved the trade-off between speed and accuracy by performing the attention operation using a DFC (decoupled fully-connected) attention. However, the DFC attention has limitations in capturing wide spatial information, which may lead to the degradation of recognition performance.
Jo Kwang Chol, a researcher at the Institute of Information Technology, has designed a network structure with low computational cost and improved performance by combining the self-attention module with the extended Ghost module based on the backbone of GhostFaceNets, and verified its accuracy using international standard databases.
The results showed that the proposed network model brings significant improvement in face recognition performance with 99.74% in LFW and 97.7% in AgeDB-30 and that with 42 MFLOP, it can support stable real-time face recognition in embedded devices.
For more details, you can refer to his paper “GhostFormerNet: A Lightweight Face Recognition Method based on Extended Ghost Module and Self-Attention” in “2025 International Conference on Graphics and Signal Processing” (EI).