EdgeFace: Efficient Face Recognition Model for Edge Devices

TL;DR

EdgeFace addresses the challenge of deploying accurate face recognition on resource-constrained edge devices by integrating CNN and Transformer strengths through a Split Depth-wise Transpose Attention (STDA) encoder and a Low Rank Linear (LoRaLin) module. The approach extends the EdgeNeXt backbone, adds a 512‑D embedding head, and trains with CosFace to obtain compact models (as low as 1.77M parameters) that achieve competitive to state-of-the-art results on standard benchmarks. Its key contributions are the LoRaLin module for efficient linear layers, the STDA encoder for global feature interactions with low cost, and extensive evaluation including the IJCB 2023 EFAR competition where EdgeFace variants rank highly across parameter budgets. The work demonstrates a practical path to high-accuracy face recognition on edge devices, with potential refinements via distillation and quantization for even greater efficiency and deployability.

Abstract

In this paper, we present EdgeFace, a lightweight and efficient face recognition network inspired by the hybrid architecture of EdgeNeXt. By effectively combining the strengths of both CNN and Transformer models, and a low rank linear layer, EdgeFace achieves excellent face recognition performance optimized for edge devices. The proposed EdgeFace network not only maintains low computational costs and compact storage, but also achieves high face recognition accuracy, making it suitable for deployment on edge devices. Extensive experiments on challenging benchmark face datasets demonstrate the effectiveness and efficiency of EdgeFace in comparison to state-of-the-art lightweight models and deep face recognition models. Our EdgeFace model with 1.77M parameters achieves state of the art results on LFW (99.73%), IJB-B (92.67%), and IJB-C (94.85%), outperforming other efficient models with larger computational complexities. The code to replicate the experiments will be made available publicly.

Figures (6)

• Figure 1: A schematic diagram of the proposed EdgeFace Face Recognition model. The image is adapted from the EdgeNeXt maaz2023edgenext model to show the additional elements added to convert it to a face recognition network. Specifically, we introduce LoRaLin layers and add a head to obtain the 512-dimensional embeddings.
• Figure 2: PyTorch class for a Low-Rank Linear layer (LoRaLin).
• Figure 3: The figure shows the reduction in Model Parameters (MPARAMS) and Multiply-Accumulate Operations (MFLOPS) as a function of Rank-ratio ($\gamma$). The dotted line represents the corresponding values from the default model employing a conventional Linear layer.
• Figure 4: Performance comparison of different models on (a) LFW, (b) IJB-B, and (c) IJB-C datasets.
• Figure 5: Ablation study of EdgeFace with respect to low rank parameter ($\gamma$). The performance is evaluated as the verification accuracy of 'XS' variant on different face recognition datasets.