TY - JOUR
TI - Mobi-FaceNeXt: Research on an efficient face recognition algorithm based on lightweight convolutional neural networks
AU - Yuan Bin
AU - Du Chang-Qing
AU - Li Zi-Tian
JN - Thermal Science
PY - 2026
VL - 30
IS - 2
SP - 1191
EP - 1201
PT - Article
AB - The advent of deep learning and convolutional neural networks, in conjunction  with the unremitting expansion and refinement of face recognition datasets,  has precipitated a substantial advancement in face recognition technology  based on convolutional neural networks. Nevertheless, in real-world  implementation scenarios, numerous disadvantages persist in the deployment  of facial recognition technology. The present study focuses on the research  of face recognition algorithms based on lightweight convolutional neural  networks. A thorough analysis of prevalent facial recognition architectures  is conducted, encompassing an examination of numerous network models. The  integration of diverse network models strengths is achieved to engineer a  lightweight network, designated as Mobi-FaceNeXt, for the purpose of facial  feature extraction. While ensuring the accuracy of face recognition, efforts  are made to minimize network parameters and computational load. This makes  the algorithm deployable on general embedded platforms and devices with  limited computing and storage resources. This has significant practical  engineering implications. In the research, a joint loss function of MagFace  Loss and Center Loss is used for training, and the processed MS-Celeb-1M  dataset is utilized to enhance the learning ability of the deep face  recognition model for facial features. Depth-wise separable convolution is  employed to reduce parameters and computations, and the algorithm is  optimized to enhance the network processing speed, thereby facilitating the  extraction of facial feature information. The experimental results  demonstrate that the Mobi-FaceNeXt model can achieve a superior level of  face recognition accuracy while maintaining a low level of network  computations and parameters. The technology in question has the capacity to  satisfy the requirements of embedded devices and to extract facial feature  information with greater efficiency. This suggests a broad range of  potential applications.
DO - 10.2298/TSCI2602191Y
ER -