University of Twente Student Theses

Login

Limited Resource Optimization for Face Recognition Neural Networks: Sub-byte quantization of MobileFaceNet using QKeras

Bunda, S.T. (2022) Limited Resource Optimization for Face Recognition Neural Networks: Sub-byte quantization of MobileFaceNet using QKeras.

This is the latest version of this item.

[img] PDF
8MB
Abstract:Face recognition is one of the most populair biometric identification systems and as such is widely used. With the growing need for digital personal data security, it is crucial to seek solutions to work on personal devices. To stimulate these developments, the computational and memory footprint of these face recognition systems should be reduced to fit on edge devices. Based on the populair MobileNetV2, MobileFaceNet is a very efficient face recognition neural network with 99.15% accuracy on the LFW dataset with a model size of only 4MB using a 32-bit representation. This work presents a method to reduce the bit length of MobileFaceNet in the form of QMobileFaceNet using sub-byte quantization. This is achieved by first identifying the most strategic use of the QKeras library enabling sub-byte dynamic fixed-point quantization. This work shows that 8-bit and 4-bit versions of QMobileFaceNet can be obtained with 98.68% and 98.63% accuracy on the LFW dataset which reduces footprint to 25% and 12.5% of the original weight respectively. Both show an accuracy loss similar to the performance described by other quantization methods applied on MobileNetV2. Using mixed-precision, an accuracy of 98.17% can be achieved whilst requiring only 10% of the original weight footprint.
Item Type:Essay (Master)
Faculty:EEMCS: Electrical Engineering, Mathematics and Computer Science
Subject:53 electrotechnology, 54 computer science
Programme:Electrical Engineering MSc (60353)
Link to this item:https://purl.utwente.nl/essays/90930
Export this item as:BibTeX
EndNote
HTML Citation
Reference Manager

 

Repository Staff Only: item control page