TY - GEN
T1 - EdgeTrim-YOLO
T2 - 4th International Conference on Computer Communication and Artificial Intelligence, CCAI 2024
AU - Xu, Jielei
AU - Pan, Feng
AU - Han, Xinheng
AU - Wang, Lingzhi
AU - Wang, Yuhe
AU - Li, Weixing
N1 - Publisher Copyright:
© 2024 IEEE.
PY - 2024
Y1 - 2024
N2 - A trim YOLO framework tailored for deployment on edge devices, named EdgeTrim-YOLO, is proposed in this study. Given the limited computing resources of edge devices, traditional YOLO frameworks often fall short of meeting the requirements for real-time performance and model efficacy. To address this issue, we conducted deep optimization and customization of the YOLO framework, introducing GhostConv, DFC Attention, and structural re-parameterization training strategies into the native backbone. These modifications significantly reduced the model's complexity and computational burden while maintaining high detection accuracy on the COCO dataset. Experimental results demonstrate that, compared to the original YOLO framework, the proposed trim YOLO framework achieved an increase in inference speed by 22.4 % on CPU (ARM), 8.2% on GPU, and 19.3% on NPU, respectively, while maintaining comparable detection performance to YOLO v5s. This provides an efficient and feasible solution for real-time object detection applications on edge devices.
AB - A trim YOLO framework tailored for deployment on edge devices, named EdgeTrim-YOLO, is proposed in this study. Given the limited computing resources of edge devices, traditional YOLO frameworks often fall short of meeting the requirements for real-time performance and model efficacy. To address this issue, we conducted deep optimization and customization of the YOLO framework, introducing GhostConv, DFC Attention, and structural re-parameterization training strategies into the native backbone. These modifications significantly reduced the model's complexity and computational burden while maintaining high detection accuracy on the COCO dataset. Experimental results demonstrate that, compared to the original YOLO framework, the proposed trim YOLO framework achieved an increase in inference speed by 22.4 % on CPU (ARM), 8.2% on GPU, and 19.3% on NPU, respectively, while maintaining comparable detection performance to YOLO v5s. This provides an efficient and feasible solution for real-time object detection applications on edge devices.
KW - edge AI
KW - model lightweighting
KW - object detection
KW - YOLO
UR - http://www.scopus.com/inward/record.url?scp=85201143219&partnerID=8YFLogxK
U2 - 10.1109/CCAI61966.2024.10602964
DO - 10.1109/CCAI61966.2024.10602964
M3 - Conference contribution
AN - SCOPUS:85201143219
T3 - 2024 4th International Conference on Computer Communication and Artificial Intelligence, CCAI 2024
SP - 113
EP - 118
BT - 2024 4th International Conference on Computer Communication and Artificial Intelligence, CCAI 2024
PB - Institute of Electrical and Electronics Engineers Inc.
Y2 - 24 May 2024 through 26 May 2024
ER -