An accurate weight binarization method for a CNN object detector using double scaling factors

Xuan Truong Nguyen; Tuan Nghia Nguyen; Hyuk Jae Lee; Hyun Kim

doi:10.1109/ICEIC49074.2020.9051090

An accurate weight binarization method for a CNN object detector using double scaling factors

Xuan Truong Nguyen, Tuan Nghia Nguyen, Hyuk Jae Lee, Hyun Kim

Dept. of Electrical and Information Engineering

Seoul National University

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

Abstract

In recent years, object detectors such as you-only-look-once (YOLO) have been intensively studied owing to applications in robotics, autonomous driving, and drones. However, memory and computation complexity is widely known as the bottlenecks in implementing YOLOv2 in hardware. A common approach is to apply weight binarization. However, the existing methods suffer from a substantial degradation in detection performance. This study proposes an accurate weight binarization method with two scaling factors. Experimental results show that the proposed method reduces the performance degradation by 32.18% while maintaining the similar memory and computation requirements as the state-of-art methods.

Original language	English
Title of host publication	2020 International Conference on Electronics, Information, and Communication, ICEIC 2020
Publisher	Institute of Electrical and Electronics Engineers Inc.
ISBN (Electronic)	9781728162898
DOIs	https://doi.org/10.1109/ICEIC49074.2020.9051090
State	Published - Jan 2020
Event	2020 International Conference on Electronics, Information, and Communication, ICEIC 2020 - Barcelona, Spain Duration: 19 Jan 2020 → 22 Jan 2020

Publication series

Name	2020 International Conference on Electronics, Information, and Communication, ICEIC 2020

Conference

Conference	2020 International Conference on Electronics, Information, and Communication, ICEIC 2020
Country/Territory	Spain
City	Barcelona
Period	19/01/20 → 22/01/20

Keywords

Binarization
Binary weight
Object detector
YOLO

Access to Document

10.1109/ICEIC49074.2020.9051090

Cite this

Nguyen, X. T., Nguyen, T. N., Lee, H. J., & Kim, H. (2020). An accurate weight binarization method for a CNN object detector using double scaling factors. In 2020 International Conference on Electronics, Information, and Communication, ICEIC 2020 Article 9051090 (2020 International Conference on Electronics, Information, and Communication, ICEIC 2020). Institute of Electrical and Electronics Engineers Inc.. https://doi.org/10.1109/ICEIC49074.2020.9051090

Nguyen, Xuan Truong ; Nguyen, Tuan Nghia ; Lee, Hyuk Jae et al. / An accurate weight binarization method for a CNN object detector using double scaling factors. 2020 International Conference on Electronics, Information, and Communication, ICEIC 2020. Institute of Electrical and Electronics Engineers Inc., 2020. (2020 International Conference on Electronics, Information, and Communication, ICEIC 2020).

@inproceedings{910a1d70c7664d809569ed66dc945504,

title = "An accurate weight binarization method for a CNN object detector using double scaling factors",

abstract = "In recent years, object detectors such as you-only-look-once (YOLO) have been intensively studied owing to applications in robotics, autonomous driving, and drones. However, memory and computation complexity is widely known as the bottlenecks in implementing YOLOv2 in hardware. A common approach is to apply weight binarization. However, the existing methods suffer from a substantial degradation in detection performance. This study proposes an accurate weight binarization method with two scaling factors. Experimental results show that the proposed method reduces the performance degradation by 32.18\% while maintaining the similar memory and computation requirements as the state-of-art methods.",

keywords = "Binarization, Binary weight, Object detector, YOLO",

author = "Nguyen, \{Xuan Truong\} and Nguyen, \{Tuan Nghia\} and Lee, \{Hyuk Jae\} and Hyun Kim",

note = "Publisher Copyright: {\textcopyright} 2020 IEEE.; 2020 International Conference on Electronics, Information, and Communication, ICEIC 2020 ; Conference date: 19-01-2020 Through 22-01-2020",

year = "2020",

month = jan,

doi = "10.1109/ICEIC49074.2020.9051090",

language = "English",

series = "2020 International Conference on Electronics, Information, and Communication, ICEIC 2020",

publisher = "Institute of Electrical and Electronics Engineers Inc.",

booktitle = "2020 International Conference on Electronics, Information, and Communication, ICEIC 2020",

}

Nguyen, XT, Nguyen, TN, Lee, HJ & Kim, H 2020, An accurate weight binarization method for a CNN object detector using double scaling factors. in 2020 International Conference on Electronics, Information, and Communication, ICEIC 2020., 9051090, 2020 International Conference on Electronics, Information, and Communication, ICEIC 2020, Institute of Electrical and Electronics Engineers Inc., 2020 International Conference on Electronics, Information, and Communication, ICEIC 2020, Barcelona, Spain, 19/01/20. https://doi.org/10.1109/ICEIC49074.2020.9051090

An accurate weight binarization method for a CNN object detector using double scaling factors. / Nguyen, Xuan Truong; Nguyen, Tuan Nghia; Lee, Hyuk Jae et al.
2020 International Conference on Electronics, Information, and Communication, ICEIC 2020. Institute of Electrical and Electronics Engineers Inc., 2020. 9051090 (2020 International Conference on Electronics, Information, and Communication, ICEIC 2020).

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

TY - GEN

T1 - An accurate weight binarization method for a CNN object detector using double scaling factors

AU - Nguyen, Xuan Truong

AU - Nguyen, Tuan Nghia

AU - Lee, Hyuk Jae

AU - Kim, Hyun

PY - 2020/1

Y1 - 2020/1

N2 - In recent years, object detectors such as you-only-look-once (YOLO) have been intensively studied owing to applications in robotics, autonomous driving, and drones. However, memory and computation complexity is widely known as the bottlenecks in implementing YOLOv2 in hardware. A common approach is to apply weight binarization. However, the existing methods suffer from a substantial degradation in detection performance. This study proposes an accurate weight binarization method with two scaling factors. Experimental results show that the proposed method reduces the performance degradation by 32.18% while maintaining the similar memory and computation requirements as the state-of-art methods.

AB - In recent years, object detectors such as you-only-look-once (YOLO) have been intensively studied owing to applications in robotics, autonomous driving, and drones. However, memory and computation complexity is widely known as the bottlenecks in implementing YOLOv2 in hardware. A common approach is to apply weight binarization. However, the existing methods suffer from a substantial degradation in detection performance. This study proposes an accurate weight binarization method with two scaling factors. Experimental results show that the proposed method reduces the performance degradation by 32.18% while maintaining the similar memory and computation requirements as the state-of-art methods.

KW - Binarization

KW - Binary weight

KW - Object detector

KW - YOLO

UR - http://www.scopus.com/inward/record.url?scp=85083513239&partnerID=8YFLogxK

U2 - 10.1109/ICEIC49074.2020.9051090

DO - 10.1109/ICEIC49074.2020.9051090

M3 - Conference contribution

AN - SCOPUS:85083513239

T3 - 2020 International Conference on Electronics, Information, and Communication, ICEIC 2020

BT - 2020 International Conference on Electronics, Information, and Communication, ICEIC 2020

PB - Institute of Electrical and Electronics Engineers Inc.

T2 - 2020 International Conference on Electronics, Information, and Communication, ICEIC 2020

Y2 - 19 January 2020 through 22 January 2020

ER -

Nguyen XT, Nguyen TN, Lee HJ, Kim H. An accurate weight binarization method for a CNN object detector using double scaling factors. In 2020 International Conference on Electronics, Information, and Communication, ICEIC 2020. Institute of Electrical and Electronics Engineers Inc. 2020. 9051090. (2020 International Conference on Electronics, Information, and Communication, ICEIC 2020). doi: 10.1109/ICEIC49074.2020.9051090

An accurate weight binarization method for a CNN object detector using double scaling factors

Abstract

Publication series

Conference

Keywords

Access to Document

Other files and links

Fingerprint

Cite this