TY - GEN
T1 - A Realtime Lightweight Detection Network Framework for UAV Identification
AU - Zeng, Guoqi
AU - Pan, Shengrui
AU - Fan, Zheng
AU - Wang, Guida
AU - Wang, Ming
AU - Yue, Huanying
N1 - Publisher Copyright:
© 2022, The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd.
PY - 2022
Y1 - 2022
N2 - Detection Neural networks are memory intensive, making them difficult to deploy on embedded system with limited hardware resources. To solve the real-time problem of airborne recognition, we introduce a lightweight detection network framework based on the YOLOv3-tiny, that work together to reduce the volume of network by 3× with accuracy increasing by 1% and detection speed reaching 15 fps. The method is based on the YOLOv3-tiny, the size of which is 33.1 MB. Its reference time and mAP are 0.085 s and 78.6%, respectively. Next, we use the structure compression and improved channel pruning strategy to reduce the network volume to 10.33 MB without affecting their accuracy. Considering that model compression will have an impact on the performance of the network, we also adopt some tricks, such as adjusting learning rate and adding attention mechanism, to further improve the identification accuracy, from 78.6% to 79.3%. These measures enable the model to run better and faster on the resource-constrained airborne embedded devices or mobile applications.
AB - Detection Neural networks are memory intensive, making them difficult to deploy on embedded system with limited hardware resources. To solve the real-time problem of airborne recognition, we introduce a lightweight detection network framework based on the YOLOv3-tiny, that work together to reduce the volume of network by 3× with accuracy increasing by 1% and detection speed reaching 15 fps. The method is based on the YOLOv3-tiny, the size of which is 33.1 MB. Its reference time and mAP are 0.085 s and 78.6%, respectively. Next, we use the structure compression and improved channel pruning strategy to reduce the network volume to 10.33 MB without affecting their accuracy. Considering that model compression will have an impact on the performance of the network, we also adopt some tricks, such as adjusting learning rate and adding attention mechanism, to further improve the identification accuracy, from 78.6% to 79.3%. These measures enable the model to run better and faster on the resource-constrained airborne embedded devices or mobile applications.
KW - Channel domain attention module
KW - Lightweight detection neural network
KW - Model compression
KW - UAV identification
KW - YOLOv3-Tiny
UR - https://www.scopus.com/pages/publications/85130940371
U2 - 10.1007/978-981-16-9492-9_263
DO - 10.1007/978-981-16-9492-9_263
M3 - 会议稿件
AN - SCOPUS:85130940371
SN - 9789811694912
T3 - Lecture Notes in Electrical Engineering
SP - 2679
EP - 2688
BT - Proceedings of 2021 International Conference on Autonomous Unmanned Systems, ICAUS 2021
A2 - Wu, Meiping
A2 - Niu, Yifeng
A2 - Gu, Mancang
A2 - Cheng, Jin
PB - Springer Science and Business Media Deutschland GmbH
T2 - International Conference on Autonomous Unmanned Systems, ICAUS 2021
Y2 - 24 September 2021 through 26 September 2021
ER -