Deploying Deep Learning (DL) based object detection (OD) models in low-end devices, such as single board computers, may lead to poor performance in terms of frames-per-second (FPS). Pruning and quantization are well-known compression techniques that can potentially lead to a reduction of the computational burden of a DL model, with a possible decrease of performance in terms of detection accuracy. Motivated by the widespread introduction of face mask mandates by many institutions during the Covid-19 pandemic, we aim at training and compressing an OD model based on YOLOv4 to recognize the presence of face masks, to be deployed on a Raspberry Pi 4. We investigate the capability of different kinds of pruning and quantization techniques of increasing the FPS with respect to the uncompressed model, while retaining the detection accuracy. We quantitatively assess the pruned and quantized models in terms of Mean Average Precision (mAP) and FPS, and show that with proper pruning and quantization, the FPS can be doubled with a moderate loss in mAP. The results provide guidelines for compression of other OD models based on YOLO.

YOLO-Based Face Mask Detection on Low-End Devices Using Pruning and Quantization

Ciro Antonio Mami;Giovanni Santacatterina;Marco Zullich
;
Felice Andrea Pellegrino
2022-01-01

Abstract

Deploying Deep Learning (DL) based object detection (OD) models in low-end devices, such as single board computers, may lead to poor performance in terms of frames-per-second (FPS). Pruning and quantization are well-known compression techniques that can potentially lead to a reduction of the computational burden of a DL model, with a possible decrease of performance in terms of detection accuracy. Motivated by the widespread introduction of face mask mandates by many institutions during the Covid-19 pandemic, we aim at training and compressing an OD model based on YOLOv4 to recognize the presence of face masks, to be deployed on a Raspberry Pi 4. We investigate the capability of different kinds of pruning and quantization techniques of increasing the FPS with respect to the uncompressed model, while retaining the detection accuracy. We quantitatively assess the pruned and quantized models in terms of Mean Average Precision (mAP) and FPS, and show that with proper pruning and quantization, the FPS can be doubled with a moderate loss in mAP. The results provide guidelines for compression of other OD models based on YOLO.
2022
978-953-233-103-5
https://ieeexplore.ieee.org/document/9803406
File in questo prodotto:
File Dimensione Formato  
2022_MIPRO_MaskDetection__Copy_for_PrePrint_.pdf

Accesso chiuso

Descrizione: © 2022 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes,creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works. Link to publisher's version: https://ieeexplore.ieee.org/document/9803406
Tipologia: Bozza finale post-referaggio (post-print)
Licenza: Digital Rights Management non definito
Dimensione 369.14 kB
Formato Adobe PDF
369.14 kB Adobe PDF   Visualizza/Apri   Richiedi una copia
YOLO-Based_Face_Mask_Detection_on_Low-End_Devices_Using_Pruning_and_Quantization.pdf

Accesso chiuso

Tipologia: Documento in Versione Editoriale
Licenza: Copyright Editore
Dimensione 906.06 kB
Formato Adobe PDF
906.06 kB Adobe PDF   Visualizza/Apri   Richiedi una copia
Pubblicazioni consigliate

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11368/3028994
Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus 6
  • ???jsp.display-item.citation.isi??? ND
social impact