Sub-byte quantization of Mobile Face Recognition Convolutional Neural Networks

Sebastian Bunda; Luuk Spreeuwers; Chris Zeinstra

doi:10.1109/BIOSIG55365.2022.9897025

Sub-byte quantization of Mobile Face Recognition Convolutional Neural Networks

Sebastian Bunda, Luuk Spreeuwers, Chris Zeinstra

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › Academic › peer-review

22 Downloads (Pure)

Abstract

Converting convolutional neural networks such as MobileNets to a full integer representation is already quite a popular method to reduce the size and computational footprint of classification networks but its effect on face recognition networks is relatively unexplored. This work presents a method to reduce the size of MobileFaceNet using sub-byte quantization of the weights and activations. It was found that 8-bit and 4-bit versions of MobileFaceNet can be obtained with 98.68% and 98.63% accuracy on the LFW dataset which reduces the footprint to 25% and 12.5% of the original weights respectively. Using mixed-precision, an accuracy of 98.17% can be achieved whilst requiring only 10% of the original weight footprint. It is expected that with a larger training dataset, higher accuracies can be achieved.

Original language	English
Title of host publication	2022 International Conference of the Biometrics Special Interest Group (BIOSIG)
Place of Publication	Piscataway, NJ
Publisher	IEEE
Pages	1-5
Number of pages	5
ISBN (Electronic)	978-1-6654-7666-9
ISBN (Print)	978-1-6654-7667-6
DOIs	https://doi.org/10.1109/BIOSIG55365.2022.9897025
Publication status	Published - 16 Sept 2022
Event	21st International Conference of the Biometrics Special Interest Group, BIOSIG 2022 - Darmstadt, Germany Duration: 14 Sept 2022 → 16 Sept 2022 Conference number: 21

Publication series

Name	International Conference of the Biometrics Special Interest Group (BIOSIG)
Publisher	IEEE
Volume	2022
ISSN (Print)	1617-5468

Conference

Conference	21st International Conference of the Biometrics Special Interest Group, BIOSIG 2022
Abbreviated title	BIOSIG 2022
Country/Territory	Germany
City	Darmstadt
Period	14/09/22 → 16/09/22

Keywords

Training
Quantization (signal)
Face recognition
Convolutional neural networks
2023 OA procedure

Access to Document

10.1109/BIOSIG55365.2022.9897025

ArticleFinal published version, 3.23 MBLicence: Taverne

Cite this

Bunda, S., Spreeuwers, L., & Zeinstra, C. (2022). Sub-byte quantization of Mobile Face Recognition Convolutional Neural Networks. In 2022 International Conference of the Biometrics Special Interest Group (BIOSIG) (pp. 1-5). Article 9897025 (International Conference of the Biometrics Special Interest Group (BIOSIG); Vol. 2022). IEEE. https://doi.org/10.1109/BIOSIG55365.2022.9897025

@inproceedings{e1bbc7023972434f8633c43fa372bccd,

title = "Sub-byte quantization of Mobile Face Recognition Convolutional Neural Networks",

abstract = "Converting convolutional neural networks such as MobileNets to a full integer representation is already quite a popular method to reduce the size and computational footprint of classification networks but its effect on face recognition networks is relatively unexplored. This work presents a method to reduce the size of MobileFaceNet using sub-byte quantization of the weights and activations. It was found that 8-bit and 4-bit versions of MobileFaceNet can be obtained with 98.68% and 98.63% accuracy on the LFW dataset which reduces the footprint to 25% and 12.5% of the original weights respectively. Using mixed-precision, an accuracy of 98.17% can be achieved whilst requiring only 10% of the original weight footprint. It is expected that with a larger training dataset, higher accuracies can be achieved.",

keywords = "Training, Quantization (signal), Face recognition, Convolutional neural networks, 2023 OA procedure",

author = "Sebastian Bunda and Luuk Spreeuwers and Chris Zeinstra",

year = "2022",

month = sep,

day = "16",

doi = "10.1109/BIOSIG55365.2022.9897025",

language = "English",

isbn = "978-1-6654-7667-6",

series = "International Conference of the Biometrics Special Interest Group (BIOSIG)",

publisher = "IEEE",

pages = "1--5",

booktitle = "2022 International Conference of the Biometrics Special Interest Group (BIOSIG)",

address = "United States",

note = "21st International Conference of the Biometrics Special Interest Group, BIOSIG 2022, BIOSIG 2022 ; Conference date: 14-09-2022 Through 16-09-2022",

}

Bunda, S , Spreeuwers, L & Zeinstra, C 2022, Sub-byte quantization of Mobile Face Recognition Convolutional Neural Networks. in 2022 International Conference of the Biometrics Special Interest Group (BIOSIG)., 9897025, International Conference of the Biometrics Special Interest Group (BIOSIG), vol. 2022, IEEE, Piscataway, NJ, pp. 1-5, 21st International Conference of the Biometrics Special Interest Group, BIOSIG 2022, Darmstadt, Hesse, Germany, 14/09/22. https://doi.org/10.1109/BIOSIG55365.2022.9897025

Sub-byte quantization of Mobile Face Recognition Convolutional Neural Networks. / Bunda, Sebastian ; Spreeuwers, Luuk ; Zeinstra, Chris.
2022 International Conference of the Biometrics Special Interest Group (BIOSIG). Piscataway, NJ: IEEE, 2022. p. 1-5 9897025 (International Conference of the Biometrics Special Interest Group (BIOSIG); Vol. 2022).

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › Academic › peer-review

TY - GEN

T1 - Sub-byte quantization of Mobile Face Recognition Convolutional Neural Networks

AU - Bunda, Sebastian

AU - Spreeuwers, Luuk

AU - Zeinstra, Chris

N1 - Conference code: 21

PY - 2022/9/16

Y1 - 2022/9/16

N2 - Converting convolutional neural networks such as MobileNets to a full integer representation is already quite a popular method to reduce the size and computational footprint of classification networks but its effect on face recognition networks is relatively unexplored. This work presents a method to reduce the size of MobileFaceNet using sub-byte quantization of the weights and activations. It was found that 8-bit and 4-bit versions of MobileFaceNet can be obtained with 98.68% and 98.63% accuracy on the LFW dataset which reduces the footprint to 25% and 12.5% of the original weights respectively. Using mixed-precision, an accuracy of 98.17% can be achieved whilst requiring only 10% of the original weight footprint. It is expected that with a larger training dataset, higher accuracies can be achieved.

AB - Converting convolutional neural networks such as MobileNets to a full integer representation is already quite a popular method to reduce the size and computational footprint of classification networks but its effect on face recognition networks is relatively unexplored. This work presents a method to reduce the size of MobileFaceNet using sub-byte quantization of the weights and activations. It was found that 8-bit and 4-bit versions of MobileFaceNet can be obtained with 98.68% and 98.63% accuracy on the LFW dataset which reduces the footprint to 25% and 12.5% of the original weights respectively. Using mixed-precision, an accuracy of 98.17% can be achieved whilst requiring only 10% of the original weight footprint. It is expected that with a larger training dataset, higher accuracies can be achieved.

KW - Training

KW - Quantization (signal)

KW - Face recognition

KW - Convolutional neural networks

KW - 2023 OA procedure

U2 - 10.1109/BIOSIG55365.2022.9897025

DO - 10.1109/BIOSIG55365.2022.9897025

M3 - Conference contribution

SN - 978-1-6654-7667-6

T3 - International Conference of the Biometrics Special Interest Group (BIOSIG)

SP - 1

EP - 5

BT - 2022 International Conference of the Biometrics Special Interest Group (BIOSIG)

PB - IEEE

CY - Piscataway, NJ

T2 - 21st International Conference of the Biometrics Special Interest Group, BIOSIG 2022

Y2 - 14 September 2022 through 16 September 2022

ER -

Sub-byte quantization of Mobile Face Recognition Convolutional Neural Networks

Abstract

Publication series

Conference

Keywords

Access to Document

Fingerprint

Cite this