This letter proposes a novel lightweight deep learning object detector named LW-YOLOv4-tiny, which incorporates the convolution block feature addition module (CBFAM). The novelty of LW-YOLOv4-tiny is the use of channel-wise convolution and element-wise addition in the CBFAM instead of utilizing the concatenation of different feature maps. The model size and computation requirement are reduced by up to 16.9 Mbytes, 5.4 billion FLOPs (BFLOPS), and 11.3 FPS, which is 31.9%, 22.8%, and 30% smaller and faster than the most recent version of YOLOv4-tiny. From the MSCOCO2017 and PASCAL VOC2012 benchmarks, LW-YOLOv4-tiny achieved 40.2% and 69.3% mAP, respectively.
Min Ho KWAK
Hankuk University of Foreign Studies
Youngwoo KIM
Korea Institute of Industrial Technology
Kangin LEE
Hyundai Motor Company
Jae Young CHOI
Hankuk University of Foreign Studies
The copyright of the original papers published on this site belongs to IEICE. Unauthorized use of the original or translated papers is prohibited. See IEICE Provisions on Copyright for details.
Copy
Min Ho KWAK, Youngwoo KIM, Kangin LEE, Jae Young CHOI, "Convolution Block Feature Addition Module (CBFAM) for Lightweight and Fast Object Detection on Non-GPU Devices" in IEICE TRANSACTIONS on Information,
vol. E106-D, no. 5, pp. 1106-1110, May 2023, doi: 10.1587/transinf.2022EDL8104.
Abstract: This letter proposes a novel lightweight deep learning object detector named LW-YOLOv4-tiny, which incorporates the convolution block feature addition module (CBFAM). The novelty of LW-YOLOv4-tiny is the use of channel-wise convolution and element-wise addition in the CBFAM instead of utilizing the concatenation of different feature maps. The model size and computation requirement are reduced by up to 16.9 Mbytes, 5.4 billion FLOPs (BFLOPS), and 11.3 FPS, which is 31.9%, 22.8%, and 30% smaller and faster than the most recent version of YOLOv4-tiny. From the MSCOCO2017 and PASCAL VOC2012 benchmarks, LW-YOLOv4-tiny achieved 40.2% and 69.3% mAP, respectively.
URL: https://global.ieice.org/en_transactions/information/10.1587/transinf.2022EDL8104/_p
Copy
@ARTICLE{e106-d_5_1106,
author={Min Ho KWAK, Youngwoo KIM, Kangin LEE, Jae Young CHOI, },
journal={IEICE TRANSACTIONS on Information},
title={Convolution Block Feature Addition Module (CBFAM) for Lightweight and Fast Object Detection on Non-GPU Devices},
year={2023},
volume={E106-D},
number={5},
pages={1106-1110},
abstract={This letter proposes a novel lightweight deep learning object detector named LW-YOLOv4-tiny, which incorporates the convolution block feature addition module (CBFAM). The novelty of LW-YOLOv4-tiny is the use of channel-wise convolution and element-wise addition in the CBFAM instead of utilizing the concatenation of different feature maps. The model size and computation requirement are reduced by up to 16.9 Mbytes, 5.4 billion FLOPs (BFLOPS), and 11.3 FPS, which is 31.9%, 22.8%, and 30% smaller and faster than the most recent version of YOLOv4-tiny. From the MSCOCO2017 and PASCAL VOC2012 benchmarks, LW-YOLOv4-tiny achieved 40.2% and 69.3% mAP, respectively.},
keywords={},
doi={10.1587/transinf.2022EDL8104},
ISSN={1745-1361},
month={May},}
Copy
TY - JOUR
TI - Convolution Block Feature Addition Module (CBFAM) for Lightweight and Fast Object Detection on Non-GPU Devices
T2 - IEICE TRANSACTIONS on Information
SP - 1106
EP - 1110
AU - Min Ho KWAK
AU - Youngwoo KIM
AU - Kangin LEE
AU - Jae Young CHOI
PY - 2023
DO - 10.1587/transinf.2022EDL8104
JO - IEICE TRANSACTIONS on Information
SN - 1745-1361
VL - E106-D
IS - 5
JA - IEICE TRANSACTIONS on Information
Y1 - May 2023
AB - This letter proposes a novel lightweight deep learning object detector named LW-YOLOv4-tiny, which incorporates the convolution block feature addition module (CBFAM). The novelty of LW-YOLOv4-tiny is the use of channel-wise convolution and element-wise addition in the CBFAM instead of utilizing the concatenation of different feature maps. The model size and computation requirement are reduced by up to 16.9 Mbytes, 5.4 billion FLOPs (BFLOPS), and 11.3 FPS, which is 31.9%, 22.8%, and 30% smaller and faster than the most recent version of YOLOv4-tiny. From the MSCOCO2017 and PASCAL VOC2012 benchmarks, LW-YOLOv4-tiny achieved 40.2% and 69.3% mAP, respectively.
ER -