Min Cui

In recent years, the rapid development of lightweight convolutional neural networks (CNNs) and lightweight vision transformers (ViTs) has led to significant progress in the field of mobile computing. However, deploying facial recognition models on low-power devices (with power consumption below 10 watts) remains challenging. To address this issue, we designed a lightweight facial recognition network specifically optimized for low-power devices—MSKFaceNet (Multi-Scale Kernels Face Network). First, we propose a novel lightweight convolutional neural network module called MSKFNet. MSKFNet adopts a bottleneck design and introduces variable kernel convolutions from VarKNet, combined with channel shuffle and structural re-parameterization techniques, establishing an efficient CNN module for embedded systems. Built upon the MSKFNet module, MSKFaceNet further integrates a lightweight SE module to enhance its feature representation capabilities. Finally, we designed a real-time facial recognition attendance system based on MSKFaceNet and developed a prototype device. Experimental results showthat MSKFaceNet, with only 0.54M parameters and 0.25 GFLOPS, achieves a recognition accuracy of 99.39% on the LFW dataset while delivering an inference speed of 10.8ms on the Jetson Nano platform. The proposed attendance system effectively and accurately performs facial recognition and attendance recording, significantly improving the efficiency and fairness of attendance management.

We revisit the design of convolutional kernels in lightweight convolutional neural networks, and inspired by the recent advances in RepLKNet, we design a Variable Kernel Convolutional Network module VarKNet, which solves the problem of the imbalance between depthwise convolution and pointwise convolution in the case of depthwise separable convolution when the network width is large, and enriches the model’s receptive field. The VarKNet module adopts a multi-branch structure during training and is re-parameterized and fused into a single-path structure during inference to maintain the strong expressive ability of the model and improve the inference speed. In order to further enhance the information exchange between channels, VarKNet adds channel shuffling in the fused branches. Built on VarKNet, we designed a large-scale face recognition network VarKFaceNet. VarKFaceNet achieved A great achievement of 99.5% accuracy on the LFW dataset with 0.7M parameters and 0.24 GFLOPS. At the same time, the measured speed on the NVIDIA Jetson Nano platform is 159 times, 4.2 times, and 2.4 times that of ResNet-50, EfficientNet, and MobileFaceNet, respectively. VarKFaceNet excels in balancing speed and accuracy and is quite suitable for embedded devices with limited resources.