PaperlayerView exact revision on arXiv

Source-linked AI summary

EdgeFace: Efficient Face Recognition Model for Edge Devices

Anjith George, Christophe Ecabert, Hatef Otroshi Shahreza, Ketan Kotwal, Sebastien Marcel

arXiv:2307.01838v2cs.CVcs.CR

TL;DR

Face recognition models with high memory and computational demands are challenging to deploy on resource-constrained devices. EdgeFace adapts a hybrid CNN–Transformer architecture and introduces low-rank linear layers; the authors report strong performance across face-recognition benchmarks while maintaining low computational complexity.

  • Problem

    State-of-the-art face-recognition models often require substantial memory and computation, challenging deployment on resource-constrained devices.

    Source: §1 · p. 1

  • Method

    EdgeFace adapts the EdgeNeXt hybrid architecture for face recognition and uses LoRaLin layers to reduce computation in linear layers.

    Source: §1 · pp. 1–2

  • Results

    EdgeFace achieves superior performance to state-of-the-art methods while maintaining low parameter counts and MAdds across the reported face-recognition benchmarks.

    Source: §6 · p. 9

  • Takeaways & Limitations

    The authors present EdgeFace as an efficient, highly accurate face-recognition model tailored for edge devices.

    Source: §6 · p. 9

Abstract

from arXiv · show

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.

1 INTRODUCTION

State-of-the-art face recognition models can be difficult to deploy on resource-constrained devices because of their memory and computational demands. EdgeFace combines CNN and Transformer capabilities in a lightweight architecture and adds LoRaLin layers to reduce linear-layer computation.

1 INTRODUCTION

  • State-of-the-art face-recognition models often require substantial memory and computation, challenging deployment on mobile, robotic, and embedded devices.Source: §1 · p. 1
  • EdgeFace combines CNN and Transformer capabilities in a lightweight face-recognition network adapted from EdgeNeXt for resource-constrained devices.Its STDA encoder processes inputs and encodes multi-scale facial features while maintaining low computational costs and compact storage.Source: §1 · p. 1
  • LoRaLin replaces a high-rank fully connected-layer matrix with two lower-rank matrices, reducing parameters and computations with minimal performance compromise.The module targets linear-layer computation.Source: §1 · pp. 1–2
  • The authors report experiments on challenging face-recognition datasets and robustness under pose variation, illumination changes, and occlusion.They state that source code will be made publicly available.Source: §1 · p. 2

2 RELATED WORK

Related work surveys lightweight neural architectures and face-recognition models developed to reduce resource demands while retaining recognition performance. It also discusses hybrid CNN–Transformer designs and the remaining attention-computation bottleneck for edge inference.

2.1 Lightweight Neural Architectures

  • Lightweight neural architectures reduce computation through depthwise separable convolutions, pointwise group convolutions, channel shuffling, and multiple kernel sizes [15],,, [20].SqueezeNet achieved state-of-the-art recognition accuracy at the time with 50× fewer parameters than contemporary models [27].Source: §2.1 · p. 2

2.2 Lightweight FR Architectures

  • Other compact designs use variable group convolutions, cheap ghost feature maps, or knowledge distillation to reduce model computation or parameters [19],,.TinyFaR uses a face generator to produce synthetic images and trains a student network to match a pretrained face-recognition model’s embeddings.Source: §2.1 · p. 2, §2.2 · p. 3

2.1 Lightweight Neural Architectures

  • Hybrid vision architectures combine CNN local processing with Transformer global interaction, but multi-head attention remains an inference-time bottleneck on edge devices, [31].Source: §2.1 · p. 2

2.2 Lightweight FR Architectures

  • Lightweight face-recognition models adapt compact architectures, including MobileFaceNets, MixFaceNets, ShiftFaceNet, and ShuffleFaceNet, with reported recognition results across varying parameter counts.MobileFaceNets report 99.55% accuracy on LFW with fewer than 1M parameters; ShuffleFaceNet variants range from 0.5M to 4.5M parameters and report over 99.20% verification accuracy on LFW.Source: §2.2 · pp. 2–3
  • Further face-recognition approaches include enhanced ConvNeXt blocks and configurable GhostFaceNets, which exploit convolutional redundancy,.GhostFaceNets can be designed with as low as 61M FLOPs with nominal reduction in recognition performance.Source: §2.2 · p. 3

3 PROPOSED EDGEFACE ARCHITECTURE

EdgeFace adapts EdgeNeXt’s hybrid CNN–Transformer architecture for face recognition on edge devices, adding low-rank linear layers and a 512-dimensional embedding head. Its design reduces linear-layer costs, and training uses selected WebFace subsets with CosFace.

3.1 EdgeFace Face Recognition Model

  • EdgeFace replaces EdgeNeXt linear layers with LoRaLin and adds an adaptive-pooling, layer-norm, and LoRaLin head that outputs a 512-dimensional representation.The model uses 112 × 112 input resolution and end-to-end training with a CosFace classification head.Source: §3.1 · p. 3

3.2 EdgeNeXt Architecture

  • EdgeNeXt combines CNNs with Transformers; its Split Depth-wise Transpose Attention encoder uses channel-wise self-attention and depth-wise convolutions to capture global and multi-scale features.Transposed query-key attention computes attention across channels to achieve linear complexity; variants include extra-extra small, extra-small, and small.Source: §3.2 · pp. 3–4

3.3 Low Rank Linear Module (LoRaLin)

  • LoRaLin factorizes each fully connected layer’s M × N weight matrix into two lower-rank matrices, reducing parameters and multiply-adds.Unlike LoRA, which leaves pretrained weights in use at inference, this design reduces model parameters while accepting a trade-off in capacity.Source: §3.3 · p. 4
  • For EdgeNeXt-XS, γ ≤0.8 reduces both parameter count and computation relative to the base model.Figure 3 plots model parameters and multiply-accumulate operations against γ, with the conventional linear-layer model shown as a dotted baseline.Source: §3.3 · p. 5, Figure 3 · p. 5

3.4 Training details

  • Training uses WebFace 12M and WebFace 4M subsets, with normalized and augmented 112 × 112 images, distributed training, AdamW, and CosFace loss.Inference removes the classification head and uses the resulting 512-dimensional embedding for comparisons.Source: §3.4 · p. 5

4 EXPERIMENTS

The experiments evaluate EdgeFace across seven face-recognition benchmarks, comparing it with lightweight models and testing low-rank settings, loss functions, and competition performance. The results report strong accuracy-efficiency trade-offs, with performance depending on the model variant and evaluation setting.

4.1 Test Datasets

  • Across seven benchmarks, LFW, CA-LFW, CP-LFW, CFP-FP, and AgeDB-30 use accuracy, while IJB-B and IJB-C use TAR at FAR 1e-4.The experiments include LFW, CA-LFW, CP-LFW, CFP-FP, AgeDB-30, IJB-B, and IJB-C.Source: §4.1 · p. 5

4.2 Comparison with SOTA

  • EdgeFace-S leads 2–5M-parameter models on CA-LFW, CP-LFW, IJB-B, and IJB-C and is most FLOP-efficient; EdgeFace-XS leads its <2M category on LFW, CP-LFW, IJB-B, and IJB-C.EdgeFace-XS is second-most efficient by FLOPs in its category; ShuffleFaceNet 0.5x has fewer FLOPs but the poorest recognition performance across datasets.Source: §4.2 · p. 5, Table 2 · p. 6

4.3.1 Ablation with varying values of γ

  • Around 20% parameter and FLOP savings come with less than 0.5% accuracy drop for EdgeFace-XS at γ = 0.6.The comparison is between the default and low-rank γ = 0.6 variants.Source: Table 3 · p. 6, §4.3.1 · p. 6
  • For γ below 0.8, EdgeFace-XS reduces parameter count, FLOPs, and model size, while recognition performance declines more sharply below γ = 0.6.Performance remains satisfactory through γ = 0.6.Source: §4.3.1 · p. 6

4.3.2 Ablation studies with loss functions

  • ArcFace performs better on LFW, CP-LFW, and CFP-FP, while CosFace leads on CA-LFW and AgeDB-30; on IJB-B and IJB-C, CosFace is slightly better at FPR 1e-4, but ArcFace is much better at FPR 1e-6.The comparison uses EdgeFace-XS with γ = 0.6, trained for 50 epochs with the same hyperparameters and batch size 512.Source: §4.3.2 · p. 6, Table 5 · p. 8

4.4 IJCB’23 Efficient Face Recognition (EFaR) competition

  • In the <2M-parameter competition category, Idiap EdgeFace-XS (γ=0.6) is the overall best-performing model and ranks first for verification accuracy, including on IJB-C.In the 2–5M category, EdgeFace-S (γ=0.5) has the highest verification accuracy across evaluated benchmarks and the highest IJB-C TAR at FAR=10−4, though it does not rank first overall.Source: §4.4 · p. 8

5 DISCUSSIONS

EdgeFace combines strong recognition results with low computational cost across benchmark datasets and in the IJCB 2023 efficient face recognition competition. Its performance extends to challenging low-quality images and to competition rankings that consider efficiency and memory.

5 DISCUSSIONS

  • EdgeFace led on four of seven benchmarks in each parameter group, while ranking most efficient in 2–5M and second-most efficient below 2M.Table 6's consolidated competition rank weights accuracy 70%, FLOPs 15%, and model size 15%.Source: §5 · p. 8, Table 6 · p. 9
  • On five higher-quality benchmarks EdgeFace was competitive; on low-quality IJB-B and IJB-C, it outperformed previous lightweight models in both parameter groups.The five higher-quality datasets are LFW, CA-LFW, CP-LFW, CFP-LFW, and AgeDB-30.Source: §5 · pp. 8–9
  • IJCB 2023 EdgeFace variants achieved best verification accuracy in both size groups; the under-2M model also placed first overall for accuracy, computation, and memory footprint.The competition results are reproduced in Table 6.Source: §5 · p. 9

6 CONCLUSIONS

EdgeFace combines convolutional and Transformer components with LoRaLin layers to deliver face recognition performance while keeping parameters and MAdds low. It ranked first overall among under-2M-parameter models in the IJCB 2023 competition, while distillation and quantization remain directions for future work.

6 CONCLUSIONS

  • EdgeFace combines convolutional and efficient self-attention-based encoders with LoRaLin layers, balancing local and global information while maintaining low parameters and MAdds.The paper presents the hybrid architecture as an efficient face-recognition design.Source: §6 · p. 9
  • The model ranked first overall among under-2M-parameter entries in IJCB 2023, and EdgeFace variants achieved best verification accuracy in both parameter groups.The competition is the IJCB 2023 Efficient Face Recognition Competition.Source: §6 · p. 9
  • Knowledge distillation may further enhance performance, while different quantization methods may improve storage and inference in future work.These are identified as potential directions, not demonstrated results.Source: §6 · p. 9

Related papers

Loading 2307.01838v2…